Building a Low-Latency Vector Search Engine for ScyllaDB
ScyllaDB Vector Search is now available. Learn about the design decisions, testing, and optimizations involved in achieving our performance goals. December 18, 2025 Update: ScyllaDB Vector Search is now GA and production-ready ScyllaDB Vector Search is now available. It brings millisecond-latency vector retrieval to massive scale. This makes ScyllaDB optimal for large-scale semantic search and retrieval-augmented generation workloads. See the Quick Start Guide and give it try Contact us with your questions, or for a personalized tour In this blog post, we share a bit about what was involved in introducing low latency and high throughput Vector Search to ScyllaDB. We’ll cover the architectural design decisions behind our integration of ScyllaDB’s shard-per-core for real-time operations and high-performance ANN processing. Additionally, we’ll look at some unexpected performance challenges we encountered and how we addressed them. If you’re really just looking for some early performance numbers, here you go: ScyllaDB Vector Search outperforms industry averages in both throughput and latency. Using public VectorDBBench datasets, it sustained up to 65K QPS (P99 < 20ms) on openai_small_50k, and 12K QPS (P99 < 40ms) on laion_large_100m. Across both configurations, tests demonstrate consistently high recall accuracy and predictable latencies, even under extreme concurrency. Why Vector Search for ScyllaDB? You might be wondering why we built Vector Search for ScyllaDB. Many vendors offer Vector Search, but we had some unique goals when we started our journey. ScyllaDB’s architecture is recognized for its performance. Users have been relying on us for real-time ML, predictive analytics, fraud detection and other latency-sensitive AI workloads for years. A growing number of users mentioned they were working with third-party Vector Search databases, but found them overly complex (and costly) to manage at scale. So we committed to building integrated low-latency vector search for ScyllaDB scale. We started with the question: How do we bring ScyllaDB’s low latencies and high throughput to something as complex as Vector Search? Most built-in vector solutions sacrifice performance for accuracy or scale. We wanted to deliver all three. Vector Search Design Decisions and Architecture Note: The topics in the remainder of this blog will be covered in more detail during P99 CONF, a free + virtual conference on all things performance. Join us live to learn more and ask questions. Rather than embedding HNSW indexing directly into the core database, we decoupled vector indexing and similarity search into a dedicated Rust engine. ScyllaDB replicas are paired with a local Vector Store node living under the same availability zone as the core ScyllaDB database. ScyllaDB nodes store tables with vectors and other data. The Vector Store service builds internal indexes based on the data read from these tables. Vector Store retrieves data from ScyllaDB using its native CQL protocol and CDC functionality. The client performs a CQL query on ScyllaDB, then ScyllaDB requests the list of neighbors from the Vector Store index using HTTP. Why did we design it this way? It allows the database and Vector Store nodes to scale independently. Running each component on its own VM lets you fine-tune hardware types: SSTables live on storage-optimized nodes, while vectors benefit from RAM-optimized ones. Traffic remains zone-local, optimizing network transfer costs for intensive workloads. It isolates the performance of regular queries in contrast to ANN queries to optimize latency. This allows real-time ingestion to progress while updates get transparently replicated to the Vector Store for inferencing. From the user’s perspective, clients simply issue ANN queries to ScyllaDB via the CQL API, and ScyllaDB transparently requests the list of neighbors from the Vector Store. The vector type is already supported by ScyllaDB’s Java, Rust, C++, Python, and C# drivers; it’s coming soon for GoCQL. Vector Store Architecture The core of our Vector Store is built on top of the USearch engine. We also use a set of Rust services to interface with ScyllaDB, build vector indexes, and provide search capabilities. The Vector Store service is built based on the Actor Framework architecture, using Rust, Tokio, Axum, and USearch. Its functionality is divided into several actors: “httpd” serves as a REST API endpoint for executing ANN searches. “db” and “db-index” are responsible for communicating with ScyllaDB. Specifically, “db-index” is responsible for building an index upfront when created (via a full table scan), as well as consuming CDC streams and forwarding those results to “monitor-items” to update the underlying index. “db” retrieves schema information and handles metadata changes (like DROP’ing an index), therefore ensuring that the underlying Vector Store remains consistent with ScyllaDB. Communication between actors is done using Tokio channels (queues) using async-await Rust features. There’s also a separate actor type for search functionality. It encapsulates all USearch computations and serves as a foundation for the entire service. One important note about our current implementation: for optimal performance, the Vector Store keeps all indexes in memory. This means that the entire index needs to fit into a single node’s RAM. We’re exploring hybrid approaches for future iterations. Building an Index We extended ScyllaDB with a CUSTOM INDEX function, as well as a set of options that the Vector Store service uses to build the index. The Vector Store service will first perform a full table scan to build the initial index. After that, the Vector Store index is kept in sync with ScyllaDB via Change Data Capture (CDC). Each write appends an entry to ScyllaDB’s CDC log, which the Vector Store service eventually consumes to keep its corresponding index consistent. A key design choice is that the Vector Store holds only the primary key and its corresponding vector embedding in memory. This greatly reduces the Vector Store memory requirements. When an ANN query runs (as shown above by the ANN OF syntax with a LIMIT clause), it will return just the list of primary keys back to the ScyllaDB caller. Those keys are then used by ScyllaDB internally to service the ResultSet back to the caller application. Testing and Optimizing Performance Update: Read our latest benchmark: 1B vectors with 2ms P99s and 250K QPS throughput Read 1B Vector Benchmark While building low-latency systems is no easy task, building low-latency Vector Stores is an even harder problem. Not surprisingly, we went through quite a few testing + optimization loops before reaching our latency targets for the Early Access program. Our basic testing environment involved a single shared instance in AWS, where we manually pinned CPUs to each process via cgroups. Next, we loaded a small dataset using VectorDBBench and proceeded with testing performance using the same set of parameters through each run. Even though we used a single instance, we decided to use a replication factor of 3 to simulate the load of a small Cloud cluster. Next, to define our embeddings, we used the ScyllaDB native vector type during table creation. We built an index as described above. Then, we microbenchmarked both CQL ANN OF queries through ScyllaDB. We also benchmarked direct requests to the in-memory Vector Store. Once done, we compared QPS and P99 latency under increasing concurrency levels to identify bottlenecks in our integration layer. Exploring the Latency Penalty of Nagle’s Algorithm Our initial benchmarks against ScyllaDB produced an unexpected result. Even at very low concurrency, we observed latencies around 50ms. More interestingly, latency remained nearly constant as we increased concurrency, indicating that the system wasn’t struggling to handle additional load. The bottleneck had to be elsewhere. When we compared ScyllaDB queries with requests sent directly to the Vector Store, the difference became clear. Vector Store queries returned in single-digit milliseconds and scaled smoothly until around 5K QPS. In contrast, ScyllaDB requests showed much higher P99 latency, which directly reduced throughput. At low concurrency, the gap between the two paths was about 46ms: a clue that pointed to a networking issue. A network capture confirmed it. Linux’s TCP Delayed ACK can wait up to 40ms before sending acknowledgments. Combined with Nagle’s algorithm, which buffers small packets until an ACK arrives, this created a feedback loop that directly inflated ScyllaDB’s latencies. The fix was straightforward: disable Nagle’s algorithm with the TCP_NODELAY socket option. With Nagle disabled, ScyllaDB latencies dropped to nearly match those of direct Vector Store queries. That said, throughput was still lower. While the Vector Store sustained ~5K QPS, ScyllaDB saturated around ~3K QPS. And that led to, of course, more testing and more optimization. Experimenting with Thread Layouts Our tests measuring performance across different thread layouts for our Vector Store service also yielded some interesting results. Each layout implements a different set of asynchronous and synchronous threads. Async threads are provided by the Rust Tokio runtime. They’re primarily used for I/O intensive computation, like networking and actor coordination. Synchronous threads used Rayon to execute CPU-intensive USearch tasks. The image below shows the layouts we implemented. The letter ‘a’ denotes a thread for asynchronous (io-intensive) computation and ‘s’ indicates a thread for synchronous (cpu-intensive) computation. For example, a1s3, stands for one asynchronous thread with three synchronous threads. The initial results below show that the layout using only asynchronous tasks provided the best QPS, at the expense of higher latency in high concurrency tests. The lowest latency was observed when threads weren’t fighting for CPU resources, with one asynchronous task and three synchronous threads. This layout, however, also provided the lowest QPS compared with all other tests. Looking at other variants (below), we can see that while oversubscribing CPUs (a1s4) does improve QPS to some extent, it comes at a significant latency cost. Dedicating one thread per CPU (a1s3) provided lower latency in contrast. Similarly, oversubscribing a single CPU for asynchronous processing also performed better than oversubscribing all CPU cores for both async and synchronous work. See those results below. Therefore, the only optimization opportunity we found here was to reduce latency on the asynchronous-only variant. The chart below shows that its latency is lower than the oversubscribed one, but grows at a faster pace under higher concurrency. So in summary, we found that: Async only (a4s0) delivered the best QPS, but latencies rose sharply at higher concurrency. Mixed (a1s3) avoided CPU contention, yielding the lowest latencies (but also the lowest QPS). Oversubscribed setups (a1s4, a4s4) gained some throughput (but at the cost of latency). The key takeaway is that adding sync threads improved latency at the cost of throughput, while async-only favored throughput but suffered under load. More Latency Optimizations A closer look at CPU traces revealed why. Each ANN request runs a burst of USearch computation. However, under concurrency, tasks preempt one another. This delays completions and hurts P99 latency. Tokio doesn’t offer task prioritization, but we implemented a neat trick: inserting a yield_now before starting USearch computation. This moved new tasks to the back of the queue, giving in-flight requests a chance to finish first. Comparing both approaches side by side (below) shows that our one-line code change provides marginally worse throughput, but big latency wins. As you can see below, the asynchronous-only, yield layout also drives even lower latency than the previous oversubscribed setup. Moreover, the graph below shows that it still drives higher QPS and now lower latencies than the mixed non-oversubscribed layout. It’s quite fascinating what a single line of code can do these days… Scaling with ScyllaDB Cloud Finally, we turned to ScyllaDB Cloud environments to test scaling. On the R7i.xlarge, we started by replicating the same tests that we ran in our previous single-node setup. Here, each ANN query retrieves the 100 most similar neighbors. This is quite a compute-intensive operation, often used for re-ranking scenarios. We achieved the same 5K QPS with single-digit millisecond latencies under moderate concurrency, while we approached the saturation point somewhere close to a concurrency of 80. Using R7i.8xlarge instances, we scaled our setup by 4X: going from 4 vCPUs to 16 vCPUs per node. Here, we ran two series of tests. For the 100 most similar neighbors, throughput saturates between 13 to 14K QPS while latency remains below 5ms under low concurrency, up to 20ms under a concurrency of 100. For the 10 most similar neighbors, throughput saturates at 20K QPS, with single-digit millisecond latencies even under a concurrency of 100. Large-Scale Performance Test Our final test involved scaling the Vector Store nodes to 64 CPUs per node. Our goal here was to get enough memory to run a larger dataset with 100M embeddings at 768 dimensions. This scale is rarely published by other vector search providers, and it still leaves plenty of headroom for even larger datasets. With 100M embeddings, we reached 12K QPS with P99 latency ranging between 20ms at low concurrency to 40ms at 200 concurrency, while maintaining over 97% recall. For comparison, the smaller dataset reached around 65K QPS for k=10 while keeping latencies steadily low even under extreme concurrency. Of course, your mileage may vary. Our tests ran on static datasets, and real-world workloads may behave differently. Still, the trajectory is promising, and we’re continuing to push towards linear scaling. Next Steps ScyllaDB Vector Search was built for users with real-time workload needs; our architecture isolates similarity function computation from the database and abstracts complexity for the user. This blog has outlined some of the design decisions, testing, and optimization involved in achieving those performance goals. We’re excited about the results of these early performance tests, and we hope you are too. We’re eager to hear our community’s feedback. Give a try, share your feedback, and help shape the future of this product. See the Quick Start Guide and give it try Contact us with your questions, or for a personalized tour
Inside the Database Internals Talks at P99 CONF 2025
“Never write a database. Even if you want to, even if you think you should. Resist. Never write a database. Unless you have to write a database. But you don’t.” – Charity MajorsBut someone has to write the databases that others rely on. Hearing about the engineering challenges they’re tackling is both fascinating and Schadenfreude-invoking – so perfect tech conference material. 😉 Since database performance is so near and dear to ScyllaDB, we reached out to our friends and colleagues across the community to that ensure a nice range of distributed data systems, approaches, and challenges would be represented at P99 CONF 2025. As you can see from our agenda, the response was overwhelming. A quick PSA for the uninitiated: P99 CONF is a free 2-day community event that’s intentionally virtual, highly interactive, and purely technical. It’s an immersion into all things performance. Distributed systems, database internals, Rust, C++, Java, Go, Wasm, Zig, Linux kernel, tracing, AI/ML & more – it’s all on the agenda. This year, you can look forward to first-hand engineering experiences from the likes of Pinterest, Clickhouse, Gemini, Arm, Rivian and VW Group Technology, Meta, Wayfair, Disney, NVIDIA, Turso, Neon, TigerBeetle, ScyllaDB, and too many others to list here. Here’s a sneak peek of the database internals talks you can look forward to at P99 CONF 2025… Join us at P99 CONF (free + virtual) Clickhouse’s C++ and Rust Journey Alexey Milovidov, Co-founder and CTO at Clickhouse Full rewrite from C++ to Rust or gradual integration with Rust libraries? For a large C++ codebase, only the latter works, but even then, there are many complications and rough edges. In my presentation, I will describe our experience integrating Rust and C++ code and some weird and unusual problems we had to overcome. Rethinking Durable Workflows and Queues: A Library-based Approach Qian Li, Co-founder at DBOS, Inc Durable workflow engines checkpoint program state to persistent storage (like a database) so that execution can always recover from where it left off. Most systems today rely on external orchestration: a centralized orchestrator and distributed workers communicating via message-passing. While this model is well-established, it’s often heavyweight, introducing substantial overhead, write amplification, and operational complexity. In this talk, we explore an alternative: a lightweight library-based durable workflow engine that embeds into application code and checkpoint state directly to the database. It handles queues and flow control through the database itself. This approach eliminates the need for a separate orchestrator, reduces network traffic, and improves performance by avoiding unnecessary writes. We’ll share our experience building DBOS, a library-based engine designed for simplicity and efficiency. We’ll discuss the architectural trade-offs, challenges in failure recovery, and key optimizations for scalability and maintainability. The Gory Details of a Full-Featured Userspace CPU Scheduler Avi Kivity, Co-founder and CTO at ScyllaDB Userspace CPU schedulers, which often accompany asynchronous I/O engines like io_uring and Linux AIO, are usually simplistic run-to-completion FIFO loops. This suffices for I/O bound applications, but for use cases that can be both CPU bound and I/O bound, this is not enough. Avi Kivity, CTO of ScyllaDB and co-maintainer of Seastar, will cover the design and implementation of the Seastar userspace CPU scheduler, which caters to more complex applications that require preemption and prioritization. The Tale of Taming TigerBeetle’s Tail Latency Tobias Ziegler, Software Engineer at Tigerbeetle In this talk, we dive into how we reduced TigerBeetle’s tail latency through algorithm engineering. ‘Algorithm engineering goes beyond studying theoretical complexity and considers how algorithms are executed efficiently on modern super-scalar CPUs. Specifically, we will look at Radix Sort and a k-way merge and explore how to implement them efficiently. We then demonstrate how we apply these algorithms incrementally to avoid latency spikes in practice. Why We’re Rewriting SQLite in Rust Glauber Costa, Co-founder and CEO at Turso Over two years ago, we forked SQLite. We were huge fans of the embedded nature of SQLite, but wanted a more open model of development…and libSQL was born as an Open Contribution project. Last year, as we were adding Vector Search to SQLite, we had a crazy idea. What could we achieve if we were to completely rewrite SQLite in Rust? This talk explains what drove us down this path, how we’re using deterministic simulation testing to ensure the reliability of the Rust rewrite, and the lessons learned (so far). I will show how a reimagining of this iconic database can lead to performance improvements of over 500x in some cases by looking at what powers it under the hood. Shared Nothing Databases at Scale Nick Van Wiggeren, CTO at PlanetScale This talk will discuss how PlanetScale scales databases in the cloud, focusing on a shared-nothing architecture that is built around expecting failure. Nick will go into how they built low-latency high-throughput systems that span multiple nodes, availability zones, and regions, while maintaining sub-millisecond response times. This starts at the storage layer and builds all the way up to micro-optimizing the load balancer, with a lot of learning at every step of the way. Reworking the Neon IO stack: Rust+tokio+io_uring+O_DIRECT Christian Schwarz, Member of Technical Staff at Databricks Neon is a serverless Postgres platform. Recently acquired by Databricks, the same technology now also powers Databricks Lakebase. In this talk, we will dive into Pageserver, the multi-tenant storage service at the heart of the architecture. We share techniques and lessons learned from reworking its IO stack to a fully asynchronous model, with direct IO against local NVMe drives; all during a period of rapid growth. Pageserver is implemented in Rust, we use the tokio async runtime for networking, and integrate it with io_uring for filesystem access. A Deep Dive into the Seastar Event Loop Pavel Emelyanov, Principal Software Engineer at ScyllaDB The core and the basis of ScyllaDB’s outstanding performance is the Seastar framework, and the core and the basis of seastar is its event loop. In this presentation, we’ll see what the loop does in great detail, analyze the limitations that it runs in and all the consequences that follow those limitations. We’ll also learn how the loop is observed by the user and various means to understand its behavior. Cost Effective, Low Latency Vector Search In Databases: A Case Study with Azure Cosmos DB Magdalen Manohar, Senior Researcher at Microsoft We’ve integrated DiskANN, a state-of-the-art vector indexing algorithm, into Azure Cosmos DB NoSQL, a state-of-the-art cloud-native operational database. Learn how we overcame the systems and algorithmic challenges of this integration to achieve <20ms query latency at the 10 million scale, while supporting scale-out to billions of vectors via automatic partitioning. Measuring Query Latency the Hard Way: An Adventure in Impractical Postgres Monitoring Simon Notley, Observability and Optimization at EnterpriseDB Sampling the session state (as exposed by pg_stat_activity) is a surprisingly powerful way to understand how your Postgres instance spends its time. It is something I can wholeheartedly recommend to any Postgres DBA that needs a lightweight way to monitor query performance in production. However, it’s a terrible way to measure query latency, fraught with complexity and weird statistical biases that could be avoided by simply using an extension built for the job, or even log analysis. But pursuing terrible ideas can be fun, so in this talk, I dive into my adventures in measuring query latency from session sampling, generate some extremely funky charts, and end up unexpectedly performing a vector similarity search. In this talk I’ll show how instead of attempting to correct the biases that plague estimates of query latency based time-domain sampling, we can instead pre-calculate the distribution of (biased) estimates based on a range of true distributions and use vector search to compare our observed distribution to these pre-calculate ones, thereby inferring the true query latency. This ‘eccentric’ method is actually surprisingly effective, and surprisingly fun. Fast and Deterministic Full Table Scans at Scale Felipe Cardeneti Mendes, Technical Director at ScyllaDB ScyllaDB’s new tablet replication algorithm replaces static vNodes with dynamic, elastic data distribution that adapts to shifting workloads. This talk discusses how tablets enable fast, predictable full table scans by keeping operations shard-local, balancing load automatically, and scaling linearly through a simple layer of indirection. Optimizing Tiered Storage for Low-Latency Real-Time Analytics Neha Pawar, Founding Engineer and Head of Data at StarTree Real-time OLAP databases usually trade performance for cost when moving from local storage to cloud object storage. This talk shows how we extended Apache Pinot to use cloud storage while still achieving sub-second P99 latencies. We’ll cover the abstraction that makes Pinot location-agnostic, strategies like pipelining, prefetching, and selective block fetches, and how to balance local and cloud storage for both cost efficiency and speed. As Fast as Possible, But Not Faster: ScyllaDB Flow Control Nadav Har’El, Distinguished Engineer at ScyllaDB Pushing requests faster than a system can handle results in rapidly growing queues. If unchecked, it risks depleting memory and system stability. This talk discusses how we engineered ScyllaDB’s flow control for high volume ingestions, allowing it to throttle over-eager clients to exactly the right pace – not so fast that we run out of memory, but also not so slow that we let available resources go to waste. Push the Database Beyond the Edge Nikita Sivukhin, Software Engineer at Turso Almost any application can benefit from having data available locally – enabling blazing-fast access and optimized write patterns. This talk will walk you through one approach to designing a full-featured sync engine, applicable across a wide range of domains, including front-end, back-end, and machine learning training. Engineering a Low-Latency Vector Search Engine for ScyllaDB Pawel Pery, Senior Software Engineer at ScyllaDB Implementing Vector Search in ScyllaDB brings challenges from low-latency to predictable performance at scale. Rather than embedding HNSW indexing directly into the core database, we decoupled vector indexing and similarity search into a dedicated Rust engine. Learn about the architectural design decisions that enabled us to combine and integrate ScyllaDB’s shard-per-core for real-time operations and high-performance ANN processing via USearch. We Told B+ Trees to Do Sorted Sets—They Nailed It (Joe Zhou, Dragonfly) Joe Zhou, Developer Advocate at DragonflyDB Sorted sets are a critical Redis data type used for leaderboards, time-series data, and priority queues. However, Redis’s skiplist-based implementation introduces significant memory overhead—averaging 37 bytes per entry on top of the essential 16 bytes for the (member, score) pair. For large sorted sets, this inefficiency can become a major bottleneck. In this talk, we’ll explore how Dragonfly reimplemented sorted sets using a B+ tree, reducing memory overhead to just 2-3 bytes per entry while improving performance. We’ll cover: Why skiplists are inefficient for large sorted sets. How B+ trees with bucketing drastically cut memory usage while maintaining O(log N) operations. Benchmark results showing 40% lower memory and better throughput vs. Redis. This optimization, now stable in Dragonfly, demonstrates how rethinking core data structures can unlock major efficiency gains. Attendees will leave with insights into: Trade-offs between skiplists and B+ trees. Real-world impact on memory and latency (P99 improvements). Lessons from implementing a custom ranking API for B+ trees. Keynote: Andy Pavlo You can also look forward to a keynote by Andy Pavlo. We’re not revealing the topic yet, but if you know Andy, you know you won’t want to miss it. Join us at P99 CONF (free + virtual)
Building a Resilient Data Platform with Write-Ahead Log at Netflix
By Prudhviraj Karumanchi, Samuel Fu, Sriram Rangarajan, Vidhya Arvind, Yun Wang, John Lu
Introduction
Netflix operates at a massive scale, serving hundreds of millions of users with diverse content and features. Behind the scenes, ensuring data consistency, reliability, and efficient operations across various services presents a continuous challenge. At the heart of many critical functions lies the concept of a Write-Ahead Log (WAL) abstraction. At Netflix scale, every challenge gets amplified. Some of the key challenges we encountered include:
- Accidental data loss and data corruption in databases
- System entropy across different datastores (e.g., writing to Cassandra and Elasticsearch)
- Handling updates to multiple partitions (e.g., building secondary indices on top of a NoSQL database)
- Data replication (in-region and across regions)
- Reliable retry mechanisms for real time data pipeline at scale
- Bulk deletes to database causing OOM on the Key-Value nodes
All the above challenges either resulted in production incidents or outages, consumed significant engineering resources, or led to bespoke solutions and technical debt. During one particular incident, a developer issued an ALTER TABLE command that led to data corruption. Fortunately, the data was fronted by a cache, so the ability to extend cache TTL quickly together with the app writing the mutations to Kafka allowed us to recover. Absent the resilience features on the application, there would have been permanent data loss. As the data platform team, we needed to provide resilience and guarantees to protect not just this application, but all the critical applications we have at Netflix.
Regarding the retry mechanisms for real time data pipelines, Netflix operates at a massive scale where failures (network errors, downstream service outages, etc.) are inevitable. We needed a reliable and scalable way to retry failed messages, without sacrificing throughput.
With these problems in mind, we decided to build a system that would solve all the aforementioned issues and continue to serve the future needs of Netflix in the online data platform space. Our Write-Ahead Log (WAL) is a distributed system that captures data changes, provides strong durability guarantees, and reliably delivers these changes to downstream consumers. This blog post dives into how Netflix is building a generic WAL solution to address common data challenges, enhance developer efficiency, and power high-leverage capabilities like secondary indices, enable cross-region replication for non-replicated storage engines, and support widely used patterns like delayed queues.
API
Our API is intentionally simple, exposing just the essential parameters. WAL has one main API endpoint, WriteToLog, abstracting away the internal implementation and ensuring that users can onboard easily.
rpc WriteToLog (WriteToLogRequest) returns (WriteToLogResponse) {...}
/**
* WAL request message
* namespace: Identifier for a particular WAL
* lifecycle: How much delay to set and original write time
* payload: Payload of the message
* target: Details of where to send the payload
*/
message WriteToLogRequest {
string namespace = 1;
Lifecycle lifecycle = 2;
bytes payload = 3;
Target target = 4;
}
/**
* WAL response message
* durable: Whether the request succeeded, failed, or unknown
* message: Reason for failure
*/
message WriteToLogResponse {
Trilean durable = 1;
string message = 2;
}
A namespace defines where and how data is stored, providing logical separation while abstracting the underlying storage systems. Each namespace can be configured to use different queues: Kafka, SQS, or combinations of multiple. Namespace also serves as a central configuration of settings, such as backoff multiplier or maximum number of retry attempts, and more. This flexibility allows our Data Platform to route different use cases to the most suitable storage system based on performance, durability, and consistency needs.
WAL can assume different personas depending on the namespace configuration.
Persona #1 (Delayed Queues)
In the example configuration below, the Product Data Systems (PDS) namespace uses SQS as the underlying message queue, enabling delayed messages. PDS uses Kafka extensively, and failures (network errors, downstream service outages, etc.) are inevitable. We needed a reliable and scalable way to retry failed messages, without sacrificing throughput. That’s when PDS started leveraging WAL for delayed messages.
"persistenceConfigurations": {
"persistenceConfiguration": [
{
"physicalStorage": {
"type": "SQS",
},
"config": {
"wal-queue": [
"dgwwal-dq-pds"
],
"wal-dlq-queue": [
"dgwwal-dlq-pds"
],
"queue.poll-interval.secs": 10,
"queue.max-messages-per-poll": 100
}
}
]
}
Persona #2 (Generic Cross-Region Replication)
Below is the namespace configuration for cross-region replication of EVCache using WAL, which replicates messages from a source region to multiple destinations. It uses Kafka under the hood.
"persistence_configurations": {
"persistence_configuration": [
{
"physical_storage": {
"type": "KAFKA"
},
"config": {
"consumer_stack": "consumer",
"context": "This is for cross region replication for evcache_foobar",
"target": {
"euwest1": "dgwwal.foobar.cluster.eu-west-1.netflix.net",
"type": "evc-replication",
"useast1": "dgwwal.foobar.cluster.us-east-1.netflix.net",
"useast2": "dgwwal.foobar.cluster.us-east-2.netflix.net",
"uswest2": "dgwwal.foobar.cluster.us-west-2.netflix.net"
},
"wal-kafka-dlq-topics": [],
"wal-kafka-topics": [
"evcache_foobar"
],
"wal.kafka.bootstrap.servers.prefix": "kafka-foobar"
}
}
]
}
Persona #3 (Handling multi-partition mutations)
Below is the namespace configuration for supporting mutateItems API in Key-Value, where multiple write requests can go to different partitions and have to be eventually consistent. A key detail in the below configuration is the presence of Kafka and durable_storage. These data stores are required to facilitate two phase commit semantics, which we will discuss in detail below.
"persistence_configurations": {
"persistence_configuration": [
{
"physical_storage": {
"type": "KAFKA"
},
"config": {
"consumer_stack": "consumer",
"contacts": "unknown",
"context": "WAL to support multi-id/namespace mutations for dgwkv.foobar",
"durable_storage": {
"namespace": "foobar_wal_type",
"shard": "walfoobar",
"type": "kv"
},
"target": {},
"wal-kafka-dlq-topics": [
"foobar_kv_multi_id-dlq"
],
"wal-kafka-topics": [
"foobar_kv_multi_id"
],
"wal.kafka.bootstrap.servers.prefix": "kaas_kafka-dgwwal_foobar7102"
}
}
]
}
An important note is that requests to WAL support at-least once semantics due to the underlying implementation.
Under the Hood
The core architecture consists of several key components working together.
Message Producer and Message Consumer separation: The message producer receives incoming messages from client applications and adds them into the queue, while the message consumer processes messages from the queue and sends them to the targets. Because of this separation, other systems can bring their own pluggable producers or consumers, depending on their use cases. WAL’s control plane allows for a pluggable model, which, depending on the use-case, allows us to switch between different message queues.
SQS and Kafka with a dead letter queue by default: Every WAL namespace has its own message queue and gets a dead letter queue (DLQ) by default, because there can be transient errors and hard errors. Application teams using Key-Value abstraction simply need to toggle a flag to enable WAL and get all this functionality without needing to understand the underlying complexity.
- Kafka-backed namespaces: handle standard message processing
- SQS-backed namespaces: support delayed queue semantics (we added custom logic to go beyond the standard defaults enforced in terms of delay, size limits, etc)
- Complex multi-partition scenarios: use queues and durable storage
Target Flexibility: The messages added to WAL are pushed to the target datastores. Targets can be Cassandra databases, Memcached caches, Kafka queues, or upstream applications. Users can specify the target via namespace configuration and in the API itself.
Deployment Model
WAL is deployed using the Data Gateway infrastructure. This means that WAL deployments automatically come with mTLS, connection management, authentication, runtime and deployment configurations out of the box.
Each data gateway abstraction (including WAL) is deployed as a shard. A shard is a physical concept describing a group of hardware instances. Each use case of WAL is usually deployed as a separate shard. For example, the Ads Events service will send requests to WAL shard A, while the Gaming Catalog service will send requests to WAL shard B, allowing for separation of concerns and avoiding noisy neighbour problems.
Each shard of WAL can have multiple namespaces. A namespace is a logical concept describing a configuration. Each request to WAL has to specify its namespace so that WAL can apply the correct configuration to the request. Each namespace has its own configuration of queues to ensure isolation per use case. If the underlying queue of a WAL namespace becomes the bottleneck of throughput, the operators can choose to add more queues on the fly by modifying the namespace configurations. The concept of shards and namespaces is shared across all Data Gateway Abstractions, including Key-Value, Counter, Timeseries, etc. The namespace configurations are stored in a globally replicated Relational SQL database to ensure availability and consistency.
Based on certain CPU and network thresholds, the Producer group and the Consumer group of each shard will (separately) automatically scale up the number of instances to ensure the service has low latency, high throughput and high availability. WAL, along with other abstractions, also uses the Netflix adaptive load shedding libraries and Envoy to automatically shed requests beyond a certain limit. WAL can be deployed to multiple regions, so each region will deploy its own group of instances.
Solving different flavors of problems with no change to the core architecture
The WAL addresses multiple data reliability challenges with no changes to the core architecture:
Data Loss Prevention: In case of database downtime, WAL can continue to hold the incoming mutations. When the database becomes available again, replay mutations back to the database. The tradeoff is eventual consistency rather than immediate consistency, and no data loss.
Generic Data Replication: For systems like EVCache (using Memcached) and RocksDB that do not support replication by default, WAL provides systematic replication (both in-region and across-region). The target can be another application, another WAL, or another queue — it’s completely pluggable through configuration.
System Entropy and Multi-Partition Solutions: Whether dealing with writes across two databases (like Cassandra and Elasticsearch) or mutations across multiple partitions in one database, the solution is the same — write to WAL first, then let the WAL consumer handle the mutations. No more asynchronous repairs needed; WAL handles retries and backoff automatically.
Data Corruption Recovery: In case of DB corruptions, restore to the last known good backup, then replay mutations from WAL omitting the offending write/mutation.
There are some major differences between using WAL and directly using Kafka/SQS. WAL is an abstraction on the underlying queues, so the underlying technology can be swapped out depending on use cases with no code changes. WAL emphasizes an easy yet effective API that saves users from complicated setups and configurations. We leverage the control plane to pivot technologies behind WAL when needed without app or client intervention.
WAL usage at Netflix
Delay Queue
The most common use case for WAL is as a Delay Queue. If an application is interested in sending a request at a certain time in the future, it can offload its requests to WAL, which guarantees that their requests will land after the specified delay.
Netflix’s Live Origin processes and delivers Netflix live stream video chunks, storing its video data in a Key-Value abstraction backed by Cassandra and EVCache. When Live Origin decides to delete certain video data after an event is completed, it issues delete requests to the Key-Value abstraction. However, the large amount of delete requests in a short burst interfere with the more important real-time read/write requests, causing performance issues in Cassandra and timeouts for the incoming live traffic. To get around this, Key-Value issues the delete requests to WAL first, with a random delay and jitter set for each delete request. WAL, after the delay, sends the delete requests back to Key-Value. Since the deletes are now a flatter curve of requests over time, Key-Value is then able to send the requests to the datastore with no issues.
Additionally, WAL is used by many services that utilize Kafka to stream events, including Ads, Gaming, Product Data Systems, etc. Whenever Kafka requests fail for any reason, the client apps will send WAL a request to retry the kafka request with a delay. This abstracts away the backoff and retry layer of Kafka for many teams, increasing developer efficiency.
Cross-Region Replication
WAL is also used for global cross-region replication. The architecture of WAL is generic and allows any datastore/applications to onboard for cross-region replication. Currently, the largest use case is EVCache, and we are working to onboard other storage engines.
EVCache is deployed by clusters of Memcached instances across multiple regions, where each cluster in each region shares the same data. Each region’s client apps will write, read, or delete data from the EVCache cluster of the same region. To ensure global consistency, the EVCache client of one region will replicate write and delete requests to all other regions. To implement this, the EVCache client that originated the request will send the request to a WAL corresponding to the EVCache cluster and region.
Since the EVCache client acts as the message producer group in this case, WAL only needs to deploy the message consumer groups. From there, the multiple message consumers are set up to each target region. They will read from the Kafka topic, and send the replicated write or delete requests to a Writer group in their target region. The Writer group will then go ahead and replicate the request to the EVCache server in the same region.
The biggest benefits of this approach, compared to our legacy architecture, is being able to migrate from multi-tenant architecture to single tenant architecture for the most latency sensitive applications. For example, Live Origin will have its own dedicated Message Consumer and Writer groups, while a less latency sensitive service can be multi-tenant. This helps us reduce the blast radius of the issues and also prevents noisy neighbor issues.
Multi-Table Mutations
WAL is used by Key-Value service to build the MutateItems API. WAL enables the API’s multi-table and multi-id mutations by implementing 2-phase commit semantics under the hood. For this discussion, we can assume that Key-Value service is backed by Cassandra, and each of its namespaces represents a certain table in a Cassandra DB.
When a Key-Value client issues a MutateItems request to Key-Value server, the request can contain multiple PutItems or DeleteItems requests. Each of those requests can go to different ids and namespaces, or Cassandra tables.
message MutateItemsRequest {
repeated MutationRequest mutations = 1;
message MutationRequest {
oneof mutation {
PutItemsRequest put = 1;
DeleteItemsRequest delete = 2;
}
}
}
The MutateItems request operates on an eventually consistent model. When the Key-Value server returns a success response, it guarantees that every operation within the MutateItemsRequest will eventually complete successfully. Individual put or delete operations may be partitioned into smaller chunks based on request size, meaning a single operation could spawn multiple chunk requests that must be processed in a specific sequence.
Two approaches exist to ensure Key-Value client requests achieve success. The synchronous approach involves client-side retries until all mutations complete. However, this method introduces significant challenges; datastores might not natively support transactions and provide no guarantees about the entire request succeeding. Additionally, when more than one replica set is involved in a request, latency occurs in unexpected ways, and the entire request chain must be retried. Also, partial failures in synchronous processing can leave the database in an inconsistent state if some mutations succeed while others fail, requiring complex rollback mechanisms or leaving data integrity compromised. The asynchronous approach was ultimately adopted to address these performance and consistency concerns.
Given Key-Value’s stateless architecture, the service cannot maintain the mutation success state or guarantee order internally. Instead, it leverages a Write-Ahead Log (WAL) to guarantee mutation completion. For each MutateItems request, Key-Value forwards individual put or delete operations to WAL as they arrive, with each operation tagged with a sequence number to preserve ordering. After transmitting all mutations, Key-Value sends a completion marker indicating the full request has been submitted.
The WAL producer receives these messages and persists the content, state, and ordering information to a durable storage. The message producer then forwards only the completion marker to the message queue. The message consumer retrieves these markers from the queue and reconstructs the complete mutation set by reading the stored state and content data, ordering operations according to their designated sequence. Failed mutations trigger re-queuing of the completion marker for subsequent retry attempts.
Closing Thoughts
Building Netflix’s generic Write-Ahead Log system has taught us several key lessons that guided our design decisions:
Pluggable Architecture is Core: The ability to support different targets, whether databases, caches, queues, or upstream applications, through configuration rather than code changes has been fundamental to WAL’s success across diverse use cases.
Leverage Existing Building Blocks: We had control plane infrastructure, Key-Value abstractions, and other components already in place. Building on top of these existing abstractions allowed us to focus on the unique challenges WAL needed to solve.
Separation of Concerns Enables Scale: By separating message processing from consumption and allowing independent scaling of each component, we can handle traffic surges and failures more gracefully.
Systems Fail — Consider Tradeoffs Carefully: WAL itself has failure modes, including traffic surges, slow consumers, and non-transient errors. We use abstractions and operational strategies like data partitioning and backpressure signals to handle these, but the tradeoffs must be understood.
Future work
- We are planning to add secondary indices in Key-Value service leveraging WAL.
- WAL can also be used by a service to guarantee sending requests to multiple datastores. For example, a database and a backup, or a database and a queue at the same time etc.
Acknowledgements
Launching WAL was a collaborative effort involving multiple teams at Netflix, and we are grateful to everyone who contributed to making this idea a reality. We would like to thank the following teams for their roles in this launch.
- Caching team — Additional thanks to Shih-Hao Yeh, Akashdeep Goel for contributing to cross region replication for KV, EVCache etc. and owning this service.
- Product Data System team — Carlos Matias Herrero, Brandon Bremen for contributing to the delay queue design and being early adopters of WAL giving valuable feedback.
- KeyValue and Composite abstractions team — Raj Ummadisetty for feedback on API design and mutateItems design discussions. Rajiv Shringi for feedback on API design.
- Kafka and Real Time Data Infrastructure teams — Nick Mahilani for feedback and inputs on integrating the WAL client into Kafka client. Sundaram Ananthanarayan for design discussions around the possibility of leveraging Flink for some of the WAL use cases.
- Joseph Lynch for providing strategic direction and organizational support for this project.
Building a Resilient Data Platform with Write-Ahead Log at Netflix was originally published in Netflix TechBlog on Medium, where people are continuing the conversation by highlighting and responding to this story.
Building easy-cass-mcp: An MCP Server for Cassandra Operations
I’ve started working on a new project that I’d like to share, easy-cass-mcp, an MCP (Model Context Protocol) server specifically designed to assist Apache Cassandra operators.
After spending over a decade optimizing Cassandra clusters in production environments, I’ve seen teams consistently struggle with how to interpret system metrics, configuration settings, schema design, and system configuration, and most importantly, how to understand how they all impact each other. While many teams have solid monitoring through JMX-based collectors, extracting and contextualizing specific operational metrics for troubleshooting or optimization can still be cumbersome. The good news is that we now have the infrastructure to make all this operational knowledge accessible through conversational AI.
easy-cass-stress Joins the Apache Cassandra Project
I’m taking a quick break from my series on Cassandra node density to share some news with the Cassandra community: easy-cass-stress has officially been donated to the Apache Software Foundation and is now part of the Apache Cassandra project ecosystem as cassandra-easy-stress.
Why This Matters
Over the past decade, I’ve worked with countless teams struggling with Cassandra performance testing and benchmarking. The reality is that stress testing distributed systems requires tools that can accurately simulate real-world workloads. Many tools make this difficult by requiring the end user to learn complex configurations and nuance. While consulting at The Last Pickle, I set out to create an easy to use tool that lets people get up and running in just a few minutes
Azure fault domains vs availability zones: Achieving zero downtime migrations
The challenges of operating production-ready enterprise systems in the cloud are ensuring applications remain up to date, secure and benefit from the latest features. This can include operating system or application version upgrades, but it is not limited to advancements in cloud provider offerings or the retirement of older ones. Recently, NetApp Instaclustr undertook a migration activity for (almost) all our Azure fault domain customers to availability zones and Basic SKU IP addresses.
Understanding Azure fault domains vs availability zones
“Azure fault domain vs availability zone” reflects a critical distinction in ensuring high availability and fault tolerance. Fault domains offer physical separation within a data center, while availability zones expand on this by distributing workloads across data centers within a region. This enhances resiliency against failures, making availability zones a clear step forward.
The need for migrating from fault domains to availability zones
NetApp Instaclustr has supported Azure as a cloud provider for our Managed open source offerings since 2016. Originally this offering was distributed across fault domains to ensure high availability using “Basic SKU public IP Addresses”, but this solution had some drawbacks when performing particular types of maintenance. Once released by Azure in several regions we extended our Azure support to availability zones which have a number of benefits including more explicit placement of additional resources, and we leveraged “Standard SKU Public IP’s” as part of this deployment.
When we introduced availability zones, we encouraged customers to provision new workloads in them. We also supported migrating workloads to availability zones, but we had not pushed existing deployments to do the migration. This was initially due to the reduced number of regions that supported availability zones.
In early 2024, we were notified that Azure would be retiring support for Basic SKU public IP addresses in September 2025. Notably, no new Basic SKU public IPs would be created after March 1, 2025. For us and our customers, this had the potential to impact cluster availability and stability – as we would be unable to add nodes, and some replacement operations would fail.
Very quickly we identified that we needed to migrate all customer deployments from Basic SKU to Standard SKU public IPs. Unfortunately, this operation involves node-level downtime as we needed to stop each individual virtual machine, detach the IP address, upgrade the IP address to the new SKU, and then reattach and start the instance. For customers who are operating their applications in line with our recommendations, node-level downtime does not have an impact on overall application availability, however it can increase strain on the remaining nodes.
Given that we needed to perform this potentially disruptive maintenance by a specific date, we decided to evaluate the migration of existing customers to Azure availability zones.
Key migration consideration for Cassandra clusters
As with any migration, we were looking at performing this with zero application downtime, minimal additional infrastructure costs, and as safe as possible. For some customers, we also needed to ensure that we do not change the contact IP addresses of the deployment, as this may require application updates from their side. We quickly worked out several ways to achieve this migration, each with its own set of pros and cons.
For our Cassandra customers, our go to method for changing cluster topology is through a data center migration. This is our zero-downtime migration method that we have completed hundreds of times, and have vast experience in executing. The benefit here is that we can be extremely confident of application uptime through the entire operation and be confident in the ability to pause and reverse the migration if issues are encountered. The major drawback to a data center migration is the increased infrastructure cost during the migration period – as you effectively need to have both your source and destination data centers running simultaneously throughout the operation. The other item of note, is that you will need to update your cluster contact points to the new data center.
For clusters running other applications, or customers who are more cost conscious, we evaluated doing a “node by node” migration from Basic SKU IP addresses in fault domains, to Standard SKU IP addresses in availability zones. This does not have any short-term increased infrastructure cost, however the upgrade from Basic SKU public IP to Standard SKU is irreversible, and different types of public IPs cannot coexist within the same fault domain. Additionally, this method comes with reduced rollback abilities. Therefore, we needed to devise a plan to minimize risks for our customers and ensure a seamless migration.
Developing a zero-downtime node-by-node migration strategy
To achieve a zero-downtime “node by node” migration, we explored several options, one of which involved building tooling to migrate the instances in the cloud provider but preserve all existing configurations. The tooling automates the migration process as follows:
- Begin with stopping the first VM in the cluster. For cluster availability, ensure that only 1 VM is stopped at any time.
- Create an OS disk snapshot and verify its success, then do the same for data disks
- Ensure all snapshots are created and generate new disks from snapshots
- Create a new network interface card (NIC) and confirm its status is green
- Create a new VM and attach the disks, confirming that the new VM is up and running
- Update the private IP address and verify the change
- The public IP SKU will then be upgraded, making sure this operation is successful
- The public IP will then be reattached to the VM
- Start the VM
Even though the disks are created from snapshots of the original disks, we encountered several discrepancies in our testing, with settings between the original VM and the new VM. For instance, certain configurations, such as caching policies, did not automatically carry over, requiring manual adjustments to align with our managed standards.
Recognizing these challenges, we decided to extend our existing node replacement mechanism to streamline our migration process. This is done so that a new instance is provisioned with a new OS disk with the same IP and application data. The new node is configured by the Instaclustr Managed Platform to be the same as the original node.
The next challenge: our existing solution is built so that the replaced node was provisioned to be the exact same as the original. However, for this operation we needed the new node to be placed in an availability zone instead of the same fault domain. This required us to extend the replacement operation so that when we triggered the replacement, the new node was placed in the desired availability zone. Once this operation completed, we had a replacement tool that ensured that the new instance was correctly provisioned in the availability zone, with a Standard SKU, and without data loss.
Now that we had two very viable options, we went back to our existing Azure customers to outline the problem space, and the operations that needed to be completed. We worked with all impacted customers on the best migration path for their specific use case or application and worked out the best time to complete the migration. Where possible, we first performed the migration on any test or QA environments before moving onto production environments.
Collaborative customer migration success
Some of our Cassandra customers opted to perform the migration using our data center migration path, however most customers opted for the node-by-node method. We successfully migrated the existing Azure fault domain clusters over to the Availability Zone that we were targeting, with only a very small number of clusters remaining. These clusters are operating in Azure regions which do not yet support availability zones, but we were able to successfully upgrade their public IP from Basic SKUs that are set for retirement to Standard SKUs.
No matter what provider you use, the pace of development in cloud computing can require significant effort to support ongoing maintenance and feature adoption to take advantage of new opportunities. For business-critical applications, being able to migrate to new infrastructure and leverage these opportunities while understanding the limitations and impact they have on other services is essential.
NetApp Instaclustr has a depth of experience in supporting business critical applications in the cloud. You can read more about another large-scale migration we completed The worlds Largest Apache Kafka and Apache Cassandra Migration or head over to our console for a free trial of the Instaclustr Managed Platform.
The post Azure fault domains vs availability zones: Achieving zero downtime migrations appeared first on Instaclustr.
Integrating support for AWS PrivateLink with Apache Cassandra® on the NetApp Instaclustr Managed Platform
Discover how NetApp Instaclustr leverages AWS PrivateLink for secure and seamless connectivity with Apache Cassandra®. This post explores the technical implementation, challenges faced, and the innovative solutions we developed to provide a robust, scalable platform for your data needs.
Last year, NetApp achieved a significant milestone by fully
integrating AWS PrivateLink support for Apache Cassandra® into the
NetApp Instaclustr Managed Platform. Read our AWS PrivateLink
support for Apache Cassandra General Availability announcement
here. Our Product Engineering team made remarkable progress in
incorporating this feature into various NetApp Instaclustr
application offerings. NetApp now offers AWS PrivateLink support as
an Enterprise Feature add-on for the Instaclustr Managed Platform
for
Cassandra,
Kafka®,
OpenSearch®,
Cadence®, and
Valkey
.
The journey to support AWS PrivateLink for Cassandra involved considerable engineering effort and numerous development cycles to create a solution tailored to the unique interaction between the Cassandra application and its client driver. After extensive development and testing, our product engineering team successfully implemented an enterprise ready solution. Read on for detailed insights into the technical implementation of our solution.
What is AWS PrivateLink?
PrivateLink is a networking solution from AWS that provides private connectivity between Virtual Private Clouds (VPCs) without exposing any traffic to the public internet. This solution is ideal for customers who require a unidirectional network connection (often due to compliance concerns), ensuring that connections can only be initiated from the source VPC to the destination VPC. Additionally, PrivateLink simplifies network management by eliminating the need to manage overlapping CIDRs between VPCs. The one-way connection allows connections to be initiated only from the source VPC to the managed cluster hosted in our platform (target VPC)—and not the other way around.
To get an idea of what major building blocks are involved in making up an end-to-end AWS PrivateLink solution for Cassandra, take a look at the following diagram—it’s a simplified representation of the infrastructure used to support a PrivateLink cluster:
In this example, we have a 3-node Cassandra cluster at the far right with one Cassandra node per Availability Zone (or AZ). Next, we have the VPC Endpoint Service and a Network Load Balancer (NLB). The Endpoint Service is essentially the AWS PrivateLink, and by design AWS needs it to be backed by an NLB–that’s pretty much what we have to manage on our side.
On the customer side, they must create a VPC Endpoint that enables them to privately connect to the AWS PrivateLink on our end; naturally, customers will also have to use a Cassandra client(s) to connect to the cluster.
AWS PrivateLink support with Instaclustr for Apache Cassandra
To incorporate AWS PrivateLink support with Instaclustr for Apache Cassandra on our platform, we came across a few technical challenges. First and foremost, the primary challenge was relatively straightforward: Cassandra clients need to talk to each individual node in a cluster.
However, the problem is that nodes in an AWS PrivateLink cluster are only assigned private IPs; that is what the nodes would announce by default when Cassandra clients attempt to discover the topology of the cluster. Cassandra clients cannot do much with the received private IPs as they cannot be used to connect to the nodes directly in an AWS PrivateLink setup.
We devised a plan of attack to get around this problem:
- Make each individual Cassandra node listen for CQL queries on unique ports.
- Configure the NLB so it can route traffic to the appropriate node based on the relevant unique port.
- Let clients implement the AddressTranslator interface from the Cassandra driver. The custom address translator will need to translate the received private IPs to one of the VPC Endpoint Elastic Network Interface (or ENI) IPs without altering the corresponding unique ports.
To understand this approach better, consider the following example:
Suppose we have a 3-node Cassandra cluster. According to the proposed approach we will need to do the followings:
- Let the nodes listen on ports 172.16.0.1:6001 (in AZ1), 172.16.0.2: 6002 (in AZ2) and 172.16.0.3: 6003 (in AZ3)
- Configure the NLB to listen on the same set of ports
- Define and associate target groups based on the port. For instance, the listener on port 6002 will be associated with a target group containing only the node that is listening on port 6002.
- As for how the custom address translator is expected to work,
let’s assume the VPC Endpoint ENI IPs are 192.168.0.1 (in AZ1),
192.168.0.2 (in AZ2) and 192.168.0.3 (in AZ3). The address
translator should translate received addresses like so:
- 172.16.0.1:6001 --> 192.168.0.1:6001 - 172.16.0.2:6002 --> 192.168.0.2:6002 - 172.16.0.3:6003 --> 192.168.0.3:6003
The proposed approach not only solves the connectivity problem but also allows for connecting to appropriate nodes based on query plans generated by load balancing policies.
Around the same time, we came up with a slightly modified approach as well: we realized the need for address translation can be mostly mitigated if we make the Cassandra nodes return the VPC Endpoint ENI IPs in the first place.
But the excitement did not last for long! Why? Because we quickly discovered a key problem: there is a limit to the number of listeners that can be added to any given AWS NLB of just 50.
While 50 is certainly a decent limit, the way we designed our solution meant we wouldn’t be able to provision a cluster with more than 50 nodes. This was quickly deemed to be an unacceptable limitation as it is not uncommon for a cluster to have more than 50 nodes; many Cassandra clusters in our fleet have hundreds of nodes. We had to abandon the idea of address translation and started thinking about alternative solution approaches.
Introducing Shotover Proxy
We were disappointed but did not lose hope. Soon after, we devised a practical solution centred around using one of our open source products: Shotover Proxy.
Shotover Proxy is used with Cassandra clusters to support AWS PrivateLink on the Instaclustr Managed Platform. What is Shotover Proxy, you ask? Shotover is a layer 7 database proxy built to allow developers, admins, DBAs, and operators to modify in-flight database requests. By managing database requests in transit, Shotover gives NetApp Instaclustr customers AWS PrivateLink’s simple and secure network setup with the many benefits of Cassandra.
Below is an updated version of the previous diagram that introduces some Shotover nodes in the mix:
As you can see, each AZ now has a dedicated Shotover proxy node.
In the above diagram, we have a 6-node Cassandra cluster. The Cassandra cluster sitting behind the Shotover nodes is an ordinary Private Network Cluster. The role of the Shotover nodes is to manage client requests to the Cassandra nodes while masking the real Cassandra nodes behind them. To the Cassandra client, the Shotover nodes appear to be Cassandra nodes, and it is only them that make up the entire cluster! This is the secret recipe for AWS PrivateLink for Instaclustr for Apache Cassandra that enabled us to get past the challenges discussed earlier.
So how is this model made to work?
Shotover can alter certain requests from—and responses to—the client. It can examine the tokens allocated to the Cassandra nodes in its own AZ (aka rack) and claim to be the owner of all those tokens. This essentially makes them appear to be an aggregation of the nodes in its own rack.
Given the purposely crafted topology and token allocation metadata, while the client directs queries to the Shotover node, the Shotover node in turn can pass them on to the appropriate Cassandra node and then transparently send responses back. It is worth noting that the Shotover nodes themselves do not store any data.
Because we only have 1 Shotover node per AZ in this design and there may be at most about 5 AZs per region, we only need that many listeners in the NLB to make this mechanism work. As such, the 50-listener limit on the NLB was no longer a problem.
The use of Shotover to manage client driver and cluster interoperability may sound straight forward to implement, but developing it was a year-long undertaking. As described above, the initial months of development were devoted to engineering CQL queries on unique ports and the AddressTranslator interface from the Cassandra driver to gracefully manage client connections to the Cassandra cluster. While this solution did successfully provide support for AWS PrivateLink with a Cassandra cluster, we knew that the 50-listener limit on the NLB was a barrier for use and wanted to provide our customers with a solution that could be used for any Cassandra cluster, regardless of node count.
The next few months of engineering were then devoted to the Proof of Concept of an alternative solution with the goal to investigate how Shotover could manage client requests for a Cassandra cluster with any number of nodes. And so, after a solution to support a cluster with any number of nodes was successfully proved, subsequent effort was then devoted to work through stability testing the new solution, the results of that engineering being the stable solution described above.
We have also conducted performance testing to evaluate the relative performance of a PrivateLink-enabled Cassandra cluster compared to its non-PrivateLink counterpart. Multiple iterations of performance testing were executed as some adjustments to Shotover were identified from test cases and resulted in the PrivateLink-enabled Cassandra cluster throughput and latency measuring near to a standard Cassandra cluster throughput and latency.
Related content: Read more about creating an AWS PrivateLink-enabled Cassandra cluster on the Instaclustr Managed Platform
The following was our experimental setup for identifying the max throughput in terms of Operations per second of a Cassandra PrivateLink cluster in comparison to a non-Cassandra PrivateLink cluster
- Baseline node size:
i3en.xlarge - Shotover Proxy node size on Cassandra Cluster:
CSO-PRD-c6gd.medium-54 - Cassandra version:
4.1.3 - Shotover Proxy version:
0.2.0 - Other configuration: Repair and backup disabled, Client Encryption disabled
Throughput results
| Operation | Operation rate with PrivateLink and Shotover | Operation rate without PrivateLink |
| Mixed-small (3 Nodes) | 16608 | 16206 |
| Mixed-small (6 Nodes) | 33585 | 33598 |
| Mixed-small (9 Nodes) | 51792 | 51798 |
Across different cluster sizes, we observed no significant difference in operation throughput between PrivateLink and non-PrivateLink configurations.
Latency results
Latency benchmarks were conducted at ~70% of the observed peak throughput (as above) to simulate realistic production traffic.
| Operation | Ops/second | Setup | Mean Latency (ms) | Median Latency (ms) | P95 Latency (ms) | P99 Latency (ms) |
| Mixed-small (3 Nodes) | 11630 | Non-PrivateLink | 9.90 | 3.2 | 53.7 | 119.4 |
| PrivateLink | 9.50 | 3.6 | 48.4 | 118.8 | ||
| Mixed-small (6 Nodes) | 23510 | Non-PrivateLink | 6 | 2.3 | 27.2 | 79.4 |
| PrivateLink | 9.10 | 3.4 | 45.4 | 104.9 | ||
| Mixed-small (9 Nodes) | 36255 | Non-PrivateLink | 5.5 | 2.4 | 21.8 | 67.6 |
| PrivateLink | 11.9 | 2.7 | 77.1 | 141.2 |
Results indicate that for lower to mid-tier throughput levels, AWS PrivateLink introduced minimal to negligible overhead. However, at higher operation rates, we observed increased latency, most notably at the p99 mark—likely due to network level factors or Shotover.
The increase in latency is expected as AWS PrivateLink introduces an additional hop to route traffic securely, which can impact latencies, particularly under heavy load. For the vast majority of applications, the observed latencies remain within acceptable ranges. However, for latency-sensitive workloads, we recommend adding more nodes (for high load cases) to help mitigate the impact of the additional network hop introduced by PrivateLink.
As with any generic benchmarking results, performance may vary depending on specific data model, workload characteristics, and environment. The results presented here are based on specific experimental setup using standard configurations and should primarily be used to compare the relative performance of PrivateLink vs. Non-PrivateLink networking under similar conditions.
Why choose AWS PrivateLink with NetApp Instaclustr?
NetApp’s commitment to innovation means you benefit from cutting-edge technology combined with ease of use. With AWS PrivateLink support on our platform, customers gain:
- Enhanced security: All traffic stays private, never touching the internet.
- Simplified networking: No need to manage complex CIDR overlaps.
- Enterprise scalability: Handles sizable clusters effortlessly.
By addressing challenges, such as the NLB listener cap and private-to-VPC IP translation, we’ve created a solution that balances efficiency, security, and scalability.
Experience PrivateLink today
The integration of AWS PrivateLink with Apache Cassandra® is now generally available with production-ready SLAs for our customers. Log in to the Console to create a Cassandra cluster with support for AWS PrivateLink with just a few clicks today. Whether you’re managing sensitive workloads or demanding performance at scale, this feature delivers unmatched value.
Want to see it in action? Book a free demo today and experience the Shotover-powered magic of AWS PrivateLink firsthand.
Resources
- Getting started: Visit the documentation to learn how to create an AWS PrivateLink-enabled Apache Cassandra cluster on the Instaclustr Managed Platform.
- Connecting clients: Already created a Cassandra cluster with AWS PrivateLink? Click here to read about how to connect Cassandra clients in one VPC to an AWS PrivateLink-enabled Cassandra cluster on the Instaclustr Platform.
- General availability announcement: For more details, read our General Availability announcement on AWS PrivateLink support for Cassandra.
The post Integrating support for AWS PrivateLink with Apache Cassandra® on the NetApp Instaclustr Managed Platform appeared first on Instaclustr.
Compaction Strategies, Performance, and Their Impact on Cassandra Node Density
This is the third post in my series on optimizing Apache Cassandra for maximum cost efficiency through increased node density. In the first post, I examined how streaming operations impact node density and laid out the groundwork for understanding why higher node density leads to significant cost savings. In the second post, I discussed how compaction throughput is critical to node density and introduced the optimizations we implemented in CASSANDRA-15452 to improve throughput on disaggregated storage like EBS.
Cassandra Compaction Throughput Performance Explained
This is the second post in my series on improving node density and lowering costs with Apache Cassandra. In the previous post, I examined how streaming performance impacts node density and operational costs. In this post, I’ll focus on compaction throughput, and a recent optimization in Cassandra 5.0.4 that significantly improves it, CASSANDRA-15452.
This post assumes some familiarity with Apache Cassandra storage engine fundamentals. The documentation has a nice section covering the storage engine if you’d like to brush up before reading this post.
CEP-24 Behind the scenes: Developing Apache Cassandra®’s password validator and generator
Introduction: The need for an Apache Cassandra® password validator and generator
Here’s the problem: while users have always had the ability to create whatever password they wanted in Cassandra–from straightforward to incredibly complex and everything in between–this ultimately created a noticeable security vulnerability.
While organizations might have internal processes for generating secure passwords that adhere to their own security policies, Cassandra itself did not have the means to enforce these standards. To make the security vulnerability worse, if a password initially met internal security guidelines, users could later downgrade their password to a less secure option simply by using “ALTER ROLE” statements.
When internal password requirements are enforced for an individual, users face the additional burden of creating compliant passwords. This inevitably involved lots of trial-and-error in attempting to create a compliant password that satisfied complex security roles.
But what if there was a way to have Cassandra automatically create passwords that meet all bespoke security requirements–but without requiring manual effort from users or system operators?
That’s why we developed CEP-24: Password validation/generation. We recognized that the complexity of secure password management could be significantly reduced (or eliminated entirely) with the right approach–and improving both security and user experience at the same time.
The Goals of CEP-24
A Cassandra Enhancement Proposal (or CEP) is a structured process for proposing, creating, and ultimately implementing new features for the Cassandra project. All CEPs are thoroughly vetted among the Cassandra community before they are officially integrated into the project.
These were the key goals we established for CEP-24:
- Introduce a way to enforce password strength upon role creation or role alteration.
- Implement a reference implementation of a password validator which adheres to a recommended password strength policy, to be used for Cassandra users out of the box.
- Emit a warning (and proceed) or just reject “create role” and “alter role” statements when the provided password does not meet a certain security level, based on user configuration of Cassandra.
- To be able to implement a custom password validator with its own policy, whatever it might be, and provide a modular/pluggable mechanism to do so.
- Provide a way for Cassandra to generate a password which would pass the subsequent validation for use by the user.
The Cassandra Password Validator and Generator builds upon an established framework in Cassandra called Guardrails, which was originally implemented under CEP-3 (more details here).
The password validator implements a custom guardrail introduced
as part of
CEP-24. A custom guardrail can validate and generate values of
arbitrary types when properly implemented. In the CEP-24 context,
the password guardrail provides
CassandraPasswordValidator by extending
ValueValidator, while passwords are generated by
CassandraPasswordGenerator by extending
ValueGenerator. Both components work with passwords as
String type values.
Password validation and generation are configured in the
cassandra.yaml file under the
password_validator section. Let’s explore the key
configuration properties available. First, the
class_name and generator_class_name
parameters specify which validator and generator classes will be
used to validate and generate passwords respectively.
Cassandra
ships CassandraPasswordValidator and CassandraPasswordGenerator out
of the box. However, if a particular enterprise decides that they
need something very custom, they are free to implement their own
validators, put it on Cassandra’s class path and reference it in
the configuration behind class_name parameter. Same for the
validator.
CEP-24 provides implementations of the validator and generator that the Cassandra team believes will satisfy the requirements of most users. These default implementations address common password security needs. However, the framework is designed with flexibility in mind, allowing organizations to implement custom validation and generation rules that align with their specific security policies and business requirements.
password_validator: # Implementation class of a validator. When not in form of FQCN, the # package name org.apache.cassandra.db.guardrails.validators is prepended. # By default, there is no validator. class_name: CassandraPasswordValidator # Implementation class of related generator which generates values which are valid when # tested against this validator. When not in form of FQCN, the # package name org.apache.cassandra.db.guardrails.generators is prepended. # By default, there is no generator. generator_class_name: CassandraPasswordGenerator
Password quality might be looked at as the number of characteristics a password satisfies. There are two levels for any password to be evaluated – warning level and failure level. Warning and failure levels nicely fit into how Guardrails act. Every guardrail has warning and failure thresholds. Based on what value a specific guardrail evaluates, it will either emit a warning to a user that its usage is discouraged (but ultimately allowed) or it will fail to be set altogether.
This same principle applies to password evaluation – each password is assessed against both warning and failure thresholds. These thresholds are determined by counting the characteristics present in the password. The system evaluates five key characteristics: the password’s overall length, the number of uppercase characters, the number of lowercase characters, the number of special characters, and the number of digits. A comprehensive password security policy can be enforced by configuring minimum requirements for each of these characteristics.
# There are four characteristics: # upper-case, lower-case, special character and digit. # If this value is set e.g. to 3, a password has to # consist of 3 out of 4 characteristics. # For example, it has to contain at least 2 upper-case characters, # 2 lower-case, and 2 digits to pass, # but it does not have to contain any special characters. # If the number of characteristics found in the password is # less than or equal to this number, it will emit a warning. characteristic_warn: 3 # If the number of characteristics found in the password is #less than or equal to this number, it will emit a failure. characteristic_fail: 2
Next, there are configuration parameters for each characteristic which count towards warning or failure:
# If the password is shorter than this value, # the validator will emit a warning. length_warn: 12 # If a password is shorter than this value, # the validator will emit a failure. length_fail: 8 # If a password does not contain at least n # upper-case characters, the validator will emit a warning. upper_case_warn: 2 # If a password does not contain at least # n upper-case characters, the validator will emit a failure. upper_case_fail: 1 # If a password does not contain at least # n lower-case characters, the validator will emit a warning. lower_case_warn: 2 # If a password does not contain at least # n lower-case characters, the validator will emit a failure. lower_case_fail: 1 # If a password does not contain at least # n digits, the validator will emit a warning. digit_warn: 2 # If a password does not contain at least # n digits, the validator will emit a failure. digit_fail: 1 # If a password does not contain at least # n special characters, the validator will emit a warning. special_warn: 2 # If a password does not contain at least # n special characters, the validator will emit a failure. special_fail: 1
It is also possible to say that illegal sequences of certain length found in a password will be forbidden:
# If a password contains illegal sequences that are at least this long, it is invalid. # Illegal sequences might be either alphabetical (form 'abcde'), # numerical (form '34567'), or US qwerty (form 'asdfg') as well # as sequences from supported character sets. # The minimum value for this property is 3, # by default it is set to 5. illegal_sequence_length: 5
Lastly, it is also possible to configure a dictionary of passwords to check against. That way, we will be checking against password dictionary attacks. It is up to the operator of a cluster to configure the password dictionary:
# Dictionary to check the passwords against. Defaults to no dictionary. # Whole dictionary is cached into memory. Use with caution with relatively big dictionaries. # Entries in a dictionary, one per line, have to be sorted per String's compareTo contract. dictionary: /path/to/dictionary/file
Now that we have gone over all the configuration parameters, let’s take a look at an example of how password validation and generation look in practice.
Consider a scenario where a Cassandra super-user (such as the default ‘cassandra’ role) attempts to create a new role named ‘alice’.
cassandra@cqlsh> CREATE ROLE alice WITH PASSWORD = 'cassandraisadatabase' AND LOGIN = true; InvalidRequest: Error from server: code=2200 [Invalid query] message="Password was not set as it violated configured password strength policy. To fix this error, the following has to be resolved: Password contains the dictionary word 'cassandraisadatabase'. You may also use 'GENERATED PASSWORD' upon role creation or alteration."
The password is not found in the dictionary, but it is not long enough. When an operator sees this, they will try to fix it by making the password longer:
cassandra@cqlsh> CREATE ROLE alice WITH PASSWORD = 'T8aum3?' AND LOGIN = true; InvalidRequest: Error from server: code=2200 [Invalid query] message="Password was not set as it violated configured password strength policy. To fix this error, the following has to be resolved: Password must be 8 or more characters in length. You may also use 'GENERATED PASSWORD' upon role creation or alteration."
The password is finally set, but it is not completely secure. It satisfies the minimum requirements but our validator identified that not all characteristics were met.
cassandra@cqlsh> CREATE ROLE alice WITH PASSWORD = 'mYAtt3mp' AND LOGIN = true; Warnings: Guardrail password violated: Password was set, however it might not be strong enough according to the configured password strength policy. To fix this warning, the following has to be resolved: Password must be 12 or more characters in length. Passwords must contain 2 or more digit characters. Password must contain 2 or more special characters. Password matches 2 of 4 character rules, but 4 are required. You may also use 'GENERATED PASSWORD' upon role creation or alteration.
The password is finally set, but it is not completely secure. It satisfies the minimum requirements but our validator identified that not all characteristics were met.
When an operator saw this, they noticed the note about the ‘GENERATED PASSWORD’ clause which will generate a password automatically without an operator needing to invent it on their own. This is a lot of times, as shown, a cumbersome process better to be left on a machine. Making it also more efficient and reliable.
cassandra@cqlsh> ALTER ROLE alice WITH GENERATED PASSWORD; generated_password ------------------ R7tb33?.mcAX
The generated password shown above will satisfy all the rules we have configured in the cassandra.yaml automatically. Every generated password will satisfy all of the rules. This is clearly an advantage over manual password generation.
When the CQL statement is executed, it will be visible in the CQLSH history (HISTORY command or in cqlsh_history file) but the password will not be logged, hence it cannot leak. It will also not appear in any auditing logs. Previously, Cassandra had to obfuscate such statements. This is not necessary anymore.
We can create a role with generated password like this:
cassandra@cqlsh> CREATE ROLE alice WITH GENERATED PASSWORD AND LOGIN = true; or by CREATE USER: cassandra@cqlsh> CREATE USER alice WITH GENERATED PASSWORD;
When a password is generated for alice (out of scope of this documentation), she can log in:
$ cqlsh -u alice -p R7tb33?.mcAX ... alice@cqlsh>
Note: It is recommended to save password to ~/.cassandra/credentials, for example:
[PlainTextAuthProvider] username = cassandra password = R7tb33?.mcAX
and by setting auth_provider in ~/.cassandra/cqlshrc
[auth_provider] module = cassandra.auth classname = PlainTextAuthProvider
It is also possible to configure password validators in such a way that a user does not see why a password failed. This is driven by configuration property for password_validator called detailed_messages. When set to false, the violations will be very brief:
alice@cqlsh> ALTER ROLE alice WITH PASSWORD = 'myattempt'; InvalidRequest: Error from server: code=2200 [Invalid query] message="Password was not set as it violated configured password strength policy. You may also use 'GENERATED PASSWORD' upon role creation or alteration."
The following command will automatically generate a new password that meets all configured security requirements.
alice@cqlsh> ALTER ROLE alice WITH GENERATED PASSWORD;
Several potential enhancements to password generation and validation could be implemented in future releases. One promising extension would be validating new passwords against previous values. This would prevent users from reusing passwords until after they’ve created a specified number of different passwords. A related enhancement could include restricting how frequently users can change their passwords, preventing rapid cycling through passwords to circumvent history-based restrictions.
These features, while valuable for comprehensive password security, were considered beyond the scope of the initial implementation and may be addressed in future updates.
Final thoughts and next steps
The Cassandra Password Validator and Generator implemented under CEP-24 represents a significant improvement in Cassandra’s security posture.
By providing robust, configurable password policies with built-in enforcement mechanisms and convenient password generation capabilities, organizations can now ensure compliance with their security standards directly at the database level. This not only strengthens overall system security but also improves the user experience by eliminating guesswork around password requirements.
As Cassandra continues to evolve as an enterprise-ready database solution, these security enhancements demonstrate a commitment to meeting the demanding security requirements of modern applications while maintaining the flexibility that makes Cassandra so powerful.
Ready to experience CEP-24 yourself? Try it out on the Instaclustr Managed Platform and spin up your first Cassandra cluster for free.
CEP-24 is just our latest contribution to open source. Check out everything else we’re working on here.
The post CEP-24 Behind the scenes: Developing Apache Cassandra®’s password validator and generator appeared first on Instaclustr.
Introduction to similarity search: Part 2–Simplifying with Apache Cassandra® 5’s new vector data type
In Part 1 of this series, we explored how you can combine Cassandra 4 and OpenSearch to perform similarity searches with word embeddings. While that approach is powerful, it requires managing two different systems.
But with the release of Cassandra 5, things become much simpler.
Cassandra 5 introduces a native VECTOR data type and built-in Vector Search capabilities, simplifying the architecture by enabling Cassandra 5 to handle storage, indexing, and querying seamlessly within a single system.
Now in Part 2, we’ll dive into how Cassandra 5 streamlines the process of working with word embeddings for similarity search. We’ll walk through how the new vector data type works, how to store and query embeddings, and how the Storage-Attached Indexing (SAI) feature enhances your ability to efficiently search through large datasets.
The power of vector search in Cassandra 5
Vector search is a game-changing feature added in Cassandra 5 that enables you to perform similarity searches directly within the database. This is especially useful for AI applications, where embeddings are used to represent data like text or images as high-dimensional vectors. The goal of vector search is to find the closest matches to these vectors, which is critical for tasks like product recommendations or image recognition.
The key to this functionality lies in embeddings: arrays of floating-point numbers that represent the similarity of objects. By storing these embeddings as vectors in Cassandra, you can use Vector Search to find connections in your data that may not be obvious through traditional queries.
How vectors work
Vectors are fixed-size sequences of non-null values, much like lists. However, in Cassandra 5, you cannot modify individual elements of a vector — you must replace the entire vector if you need to update it. This makes vectors ideal for storing embeddings, where you need to work with the whole data structure at once.
When working with embeddings, you’ll typically store them as vectors of floating-point numbers to represent the semantic meaning.
Storage-Attached Indexing (SAI): The engine behind vector search
Vector Search in Cassandra 5 is powered by Storage-Attached Indexing, which enables high-performance indexing and querying of vector data. SAI is essential for Vector Search, providing the ability to create column-level indexes on vector data types. This ensures that your vector queries are both fast and scalable, even with large datasets.
SAI isn’t just limited to vectors—it also indexes other types of data, making it a versatile tool for boosting the performance of your queries across the board.
Example: Performing similarity search with Cassandra 5’s vector data type
Now that we’ve introduced the new vector data type and the power of Vector Search in Cassandra 5, let’s dive into a practical example. In this section, we’ll show how to set up a table to store embeddings, insert data, and perform similarity searches directly within Cassandra.
Step 1: Setting up the embeddings table
To get started with this example, you’ll need access to a Cassandra 5 cluster. Cassandra 5 introduces native support for vector data types and Vector Search, available on Instaclustr’s managed platform. Once you have your cluster up and running, the first step is to create a table to store the embeddings. We’ll also create an index on the vector column to optimize similarity searches using SAI.
CREATE KEYSPACE aisearch WITH REPLICATION = {{'class': 'SimpleStrategy', ' replication_factor': 1}}; CREATE TABLE IF NOT EXISTS embeddings ( id UUID, paragraph_uuid UUID, filename TEXT, embeddings vector<float, 300>, text TEXT, last_updated timestamp, PRIMARY KEY (id, paragraph_uuid) ); CREATE INDEX IF NOT EXISTS ann_index ON embeddings(embeddings) USING 'sai';
This setup allows us to store the embeddings as 300-dimensional vectors, along with metadata like file names and text. The SAI index will be used to speed up similarity searches on the embedding’s column.
You can also fine-tune the index by specifying the similarity function to be used for vector comparisons. Cassandra 5 supports three types of similarity functions: DOT_PRODUCT, COSINE, and EUCLIDEAN. By default, the similarity function is set to COSINE, but you can specify your preferred method when creating the index:
CREATE INDEX IF NOT EXISTS ann_index ON embeddings(embeddings) USING 'sai' WITH OPTIONS = { 'similarity_function': 'DOT_PRODUCT' };
Each similarity function has its own advantages depending on your use case. DOT_PRODUCT is often used when you need to measure the direction and magnitude of vectors, COSINE is ideal for comparing the angle between vectors, and EUCLIDEAN calculates the straight-line distance between vectors. By selecting the appropriate function, you can optimize your search results to better match the needs of your application.
Step 2: Inserting embeddings into Cassandra 5
To insert embeddings into Cassandra 5, we can use the same code from the first part of this series to extract text from files, load the FastText model, and generate the embeddings. Once the embeddings are generated, the following function will insert them into Cassandra:
import time from uuid import uuid4, UUID from cassandra.cluster import Cluster from cassandra.query import SimpleStatement from cassandra.policies import DCAwareRoundRobinPolicy from cassandra.auth import PlainTextAuthProvider from google.colab import userdata # Connect to the single-node cluster cluster = Cluster( # Replace with your IP list ["xxx.xxx.xxx.xxx", "xxx.xxx.xxx.xxx ", " xxx.xxx.xxx.xxx "], # Single-node cluster address load_balancing_policy=DCAwareRoundRobinPolicy(local_dc='AWS_VPC_US_EAST_1'), # Update the local data centre if needed port=9042, auth_provider=PlainTextAuthProvider ( username='iccassandra', password='replace_with_your_password' ) ) session = cluster.connect() print('Connected to cluster %s' % cluster.metadata.cluster_name) def insert_embedding_to_cassandra(session, embedding, id=None, paragraph_uuid=None, filename=None, text=None, keyspace_name=None): try: embeddings = list(map(float, embedding)) # Generate UUIDs if not provided if id is None: id = uuid4() if paragraph_uuid is None: paragraph_uuid = uuid4() # Ensure id and paragraph_uuid are UUID objects if isinstance(id, str): id = UUID(id) if isinstance(paragraph_uuid, str): paragraph_uuid = UUID(paragraph_uuid) # Create the query string with placeholders insert_query = f""" INSERT INTO {keyspace_name}.embeddings (id, paragraph_uuid, filename, embeddings, text, last_updated) VALUES (?, ?, ?, ?, ?, toTimestamp(now())) """ # Create a prepared statement with the query prepared = session.prepare(insert_query) # Execute the query session.execute(prepared.bind((id, paragraph_uuid, filename, embeddings, text))) return None # Successful insertion except Exception as e: error_message = f"Failed to execute query:\nError: {str(e)}" return error_message # Return error message on failure def insert_with_retry(session, embedding, id=None, paragraph_uuid=None, filename=None, text=None, keyspace_name=None, max_retries=3, retry_delay_seconds=1): retry_count = 0 while retry_count < max_retries: result = insert_embedding_to_cassandra(session, embedding, id, paragraph_uuid, filename, text, keyspace_name) if result is None: return True # Successful insertion else: retry_count += 1 print(f"Insertion failed on attempt {retry_count} with error: {result}") if retry_count < max_retries: time.sleep(retry_delay_seconds) # Delay before the next retry return False # Failed after max_retries # Replace the file path pointing to the desired file file_path = "/path/to/Cassandra-Best-Practices.pdf" paragraphs_with_embeddings = extract_text_with_page_number_and_embeddings(file_path) from tqdm import tqdm for paragraph in tqdm(paragraphs_with_embeddings, desc="Inserting paragraphs"): if not insert_with_retry( session=session, embedding=paragraph['embedding'], id=paragraph['uuid'], paragraph_uuid=paragraph['paragraph_uuid'], text=paragraph['text'], filename=paragraph['filename'], keyspace_name=keyspace_name, max_retries=3, retry_delay_seconds=1 ): # Display an error message if insertion fails tqdm.write(f"Insertion failed after maximum retries for UUID {paragraph['uuid']}: {paragraph['text'][:50]}...")
This function handles inserting embeddings and metadata into Cassandra, ensuring that UUIDs are correctly generated for each entry.
Step 3: Performing similarity searches in Cassandra 5
Once the embeddings are stored, we can perform similarity searches directly within Cassandra using the following function:
import numpy as np # ------------------ Embedding Functions ------------------ def text_to_vector(text): """Convert a text chunk into a vector using the FastText model.""" words = text.split() vectors = [fasttext_model[word] for word in words if word in fasttext_model.key_to_index] return np.mean(vectors, axis=0) if vectors else np.zeros(fasttext_model.vector_size) def find_similar_texts_cassandra(session, input_text, keyspace_name=None, top_k=5): # Convert the input text to an embedding input_embedding = text_to_vector(input_text) input_embedding_str = ', '.join(map(str, input_embedding.tolist())) # Adjusted query without the ORDER BY clause and correct comment syntax query = f""" SELECT text, filename, similarity_cosine(embeddings, ?) AS similarity FROM {keyspace_name}.embeddings ORDER BY embeddings ANN OF [{input_embedding_str}] LIMIT {top_k}; """ prepared = session.prepare(query) bound = prepared.bind((input_embedding,)) rows = session.execute(bound) # Sort the results by similarity in Python similar_texts = sorted([(row.similarity, row.filename, row.text) for row in rows], key=lambda x: x[0], reverse=True) return similar_texts[:top_k] from IPython.display import display, HTML # The word you want to find similarities for input_text = "place" # Call the function to find similar texts in the Cassandra database similar_texts = find_similar_texts_cassandra(session, input_text, keyspace_name="aisearch", top_k=10)
This function searches for similar embeddings in Cassandra and retrieves the top results based on cosine similarity. Under the hood, Cassandra’s vector search uses Hierarchical Navigable Small Worlds (HNSW). HNSW organizes data points in a multi-layer graph structure, making queries significantly faster by narrowing down the search space efficiently—particularly important when handling large datasets.
Step 4: Displaying the results
To display the results in a readable format, we can loop through the similar texts and present them along with their similarity scores:
# Print the similar texts along with their similarity scores for similarity, filename, text in similar_texts: html_content = f""" <div style="margin-bottom: 10px;"> <p><b>Similarity:</b> {similarity:.4f}</p> <p><b>Text:</b> {text}</p> <p><b>File:</b> {filename}</p> </div> <hr/> """ display(HTML(html_content))
This code will display the top similar texts, along with their similarity scores and associated file names.
Cassandra 5 vs. Cassandra 4 + OpenSearch®
Cassandra 4 relies on an integration with OpenSearch to handle word embeddings and similarity searches. This approach works well for applications that are already using or comfortable with OpenSearch, but it does introduce additional complexity with the need to maintain two systems.
Cassandra 5, on the other hand, brings vector support directly into the database. With its native VECTOR data type and similarity search functions, it simplifies your architecture and improves performance, making it an ideal solution for applications that require embedding-based searches at scale.
| Feature | Cassandra 4 + OpenSearch | Cassandra 5 (Preview) |
| Embedding Storage | OpenSearch | Native VECTOR Data Type |
| Similarity Search | KNN Plugin in OpenSearch | COSINE, EUCLIDEAN, DOT_PRODUCT |
| Search Method | Exact K-Nearest Neighbor | Approximate Nearest Neighbor (ANN) |
| System Complexity | Requires two systems | All-in-one Cassandra solution |
Conclusion: A simpler path to similarity search with Cassandra 5
With Cassandra 5, the complexity of setting up and managing a separate search system for word embeddings is gone. The new vector data type and Vector Search capabilities allow you to perform similarity searches directly within Cassandra, simplifying your architecture and making it easier to build AI-powered applications.
Coming up: more in-depth examples and use cases that demonstrate how to take full advantage of these new features in Cassandra 5 in future blogs!
Ready to experience vector search with Cassandra 5? Spin up your first cluster for free on the Instaclustr Managed Platform and try it out!
The post Introduction to similarity search: Part 2–Simplifying with Apache Cassandra® 5’s new vector data type appeared first on Instaclustr.
How Cassandra Streaming, Performance, Node Density, and Cost are All related
This is the first post of several I have planned on optimizing Apache Cassandra for maximum cost efficiency. I’ve spent over a decade working with Cassandra and have spent tens of thousands of hours data modeling, fixing issues, writing tools for it, and analyzing it’s performance. I’ve always been fascinated by database performance tuning, even before Cassandra.
A decade ago I filed one of my first issues with the project, where I laid out my target goal of 20TB of data per node. This wasn’t possible for most workloads at the time, but I’ve kept this target in my sights.
Cassandra 5 Released! What's New and How to Try it
Apache Cassandra 5.0 has officially landed! This highly anticipated release brings a range of new features and performance improvements to one of the most popular NoSQL databases in the world. Having recently hosted a webinar covering the major features of Cassandra 5.0, I’m excited to give a brief overview of the key updates and show you how to easily get hands-on with the latest release using easy-cass-lab.
You can grab the latest release on the Cassandra download page.
easy-cass-lab v5 released
I’ve got some fun news to start the week off for users of easy-cass-lab: I’ve just released version 5. There are a number of nice improvements and bug fixes in here that should make it more enjoyable, more useful, and lay groundwork for some future enhancements.
- When the cluster starts, we wait for the storage service to
reach NORMAL state, then move to the next node. This is in contrast
to the previous behavior where we waited for 2 minutes after
starting a node. This queries JMX directly using Swiss Java Knife
and is more reliable than the 2-minute method. Please see
packer/bin-cassandra/wait-for-up-normalto read through the implementation. - Trunk now works correctly. Unfortunately, AxonOps doesn’t support trunk (5.1) yet, and using the agent was causing a startup error. You can test trunk out, but for now the AxonOps integration is disabled.
- Added a new repl mode. This saves keystrokes and provides some
auto-complete functionality and keeps SSH connections open. If
you’re going to do a lot of work with ECL this will help you be a
little more efficient. You can try this out with
ecl repl. - Power user feature: Initial support for profiles in AWS regions
other than
us-west-2. We only provide AMIs forus-west-2, but you can now set up a profile in an alternate region, and build the required AMIs usingeasy-cass-lab build-image. This feature is still under development and requires using aneasy-cass-labbuild from source. Credit to Jordan West for contributing this work. - Power user feature: Support for multiple profiles. Setting the
EASY_CASS_LAB_PROFILEenvironment variable allows you to configure alternate profiles. This is handy if you want to use multiple regions or have multiple organizations. - The project now uses Kotlin instead of Groovy for Gradle configuration.
- Updated Gradle to 8.9.
- When using the list command, don’t show the alias “current”.
- Project cleanup, remove old unused pssh, cassandra build, and async profiler subprojects.
The release has been released to the project’s GitHub page and to homebrew. The project is largely driven by my own consulting needs and for my training. If you’re looking to have some features prioritized please reach out, and we can discuss a consulting engagement.
easy-cass-lab updated with Cassandra 5.0 RC-1 Support
I’m excited to announce that the latest version of easy-cass-lab now supports Cassandra 5.0 RC-1, which was just made available last week! This update marks a significant milestone, providing users with the ability to test and experiment with the newest Cassandra 5.0 features in a simplified manner. This post will walk you through how to set up a cluster, SSH in, and run your first stress test.
For those new to easy-cass-lab, it’s a tool designed to streamline the setup and management of Cassandra clusters in AWS, making it accessible for both new and experienced users. Whether you’re running tests, developing new features, or just exploring Cassandra, easy-cass-lab is your go-to tool.
easy-cass-lab now available in Homebrew
I’m happy to share some exciting news for all Cassandra enthusiasts! My open source project, easy-cass-lab, is now installable via a homebrew tap. This powerful tool is designed to make testing any major version of Cassandra (or even builds that haven’t been released yet) a breeze, using AWS. A big thank-you to Jordan West who took the time to make this happen!
What is easy-cass-lab?
easy-cass-lab is a versatile testing tool for Apache Cassandra. Whether you’re dealing with the latest stable releases or experimenting with unreleased builds, easy-cass-lab provides a seamless way to test and validate your applications. With easy-cass-lab, you can ensure compatibility and performance across different Cassandra versions, making it an essential tool for developers and system administrators. easy-cass-lab is used extensively for my consulting engagements, my training program, and to evaluate performance patches destined for open source Cassandra. Here are a few examples:
Cassandra Training Signups For July and August Are Open!
I’m pleased to announce that I’ve opened training signups for Operator Excellence to the public for July and August. If you’re interested in stepping up your game as a Cassandra operator, this course is for you. Head over to the training page to find out more and sign up for the course.
Streaming My Sessions With Cassandra 5.0
As a long time participant with the Cassandra project, I’ve witnessed firsthand the evolution of this incredible database. From its early days to the present, our journey has been marked by continuous innovation, challenges, and a relentless pursuit of excellence. I’m thrilled to share that I’ll be streaming several working sessions over the next several weeks as I evaluate the latest builds and test out new features as we move toward the 5.0 release.
Streaming Cassandra Workloads and Experiments
Streaming
In the world of software engineering, especially within the realm of distributed systems, continuous learning and experimentation are not just beneficial; they’re essential. As a software engineer with a focus on distributed systems, particularly Apache Cassandra, I’ve taken this ethos to heart. My journey has led me to not only explore the intricacies of Cassandra’s distributed architecture but also to share my experiences and findings with a broader audience. This is why my YouTube channel has become an active platform where I stream at least once a week, engaging with viewers through coding sessions, trying new approaches, and benchmarking different Cassandra workloads.
Live Streaming On Tuesdays
As I promised in December, I redid my presentation from the Cassandra Summit 2023 on a live stream. You can check it out at the bottom of this post.
Going forward, I’ll be live-streaming on Tuesdays at 10AM Pacific on my YouTube channel.
Next week I’ll be taking a look at tlp-stress, which is used by the teams at some of the biggest Cassandra deployments in the world to benchmark their clusters. You can find that here.
Cassandra Summit Recap: Performance Tuning and Cassandra Training
Hello, friends in the Apache Cassandra community!
I recently had the pleasure of speaking at the Cassandra Summit in San Jose. Unfortunately, we ran into an issue with my screen refusing to cooperate with the projector, so my slides were pretty distorted and hard to read. While the talk is online, I think it would be better to have a version with the right slides as well as a little more time. I’ve decided to redo the entire talk via a live stream on YouTube. I’m scheduling this for 10am PST on Wednesday, January 17 on my YouTube channel. My original talk was done in 30 minute slot, this will be a full hour, giving plenty of time for Q&A.
Cassandra Summit, YouTube, and a Mailing List
I am thrilled to share some significant updates and exciting plans with my readers and the Cassandra community. As we draw closer to the end of the year, I’m preparing for an important speaking engagement and mapping out a year ahead filled with engaging and informative activities.
Cassandra Summit Presentation: Mastering Performance Tuning
I am honored to announce that I will be speaking at the upcoming Cassandra Summit. My talk, titled “Cassandra Performance Tuning Like You’ve Been Doing It for Ten Years,” is scheduled for December 13th, from 4:10 pm to 4:40 pm. This session aims to equip attendees with advanced insights and practical skills for optimizing Cassandra’s performance, drawing from a decade’s worth of experience in the field. Whether you’re new to Cassandra or a seasoned user, this talk will provide valuable insights to enhance your database management skills.
Uncover Cassandra's Throughput Boundaries with the New Adaptive Scheduler in tlp-stress
Introduction
Apache Cassandra remains the preferred choice for organizations seeking a massively scalable NoSQL database. To guarantee predictable performance, Cassandra administrators and developers rely on benchmarking tools like tlp-stress, nosqlbench, and ndbench to help them discover their cluster’s limits. In this post, we will explore the latest advancements in tlp-stress, highlighting the introduction of the new Adaptive Scheduler. This brand-new feature allows users to more easily uncover the throughput boundaries of Cassandra clusters while remaining within specific read and write latency targets. First though, we’ll take a brief look at the new workload designed to stress test the new Storage Attached Indexes feature coming in Cassandra 5.
AxonOps Review - An Operations Platform for Apache Cassandra
Note: Before we dive into this review of AxonOps and their offerings, it’s important to note that this blog post is part of a paid engagement in which I provided product feedback. AxonOps had no influence or say over the content of this post and did not have access to it prior to publishing.
In the ever-evolving landscape of data management, companies are constantly seeking solutions that can simplify the complexities of database operations. One such player in the market is AxonOps, a company that specializes in providing tooling for operating Apache Cassandra.
Benchmarking Apache Cassandra with tlp-stress
This post will introduce you to tlp-stress, a tool for benchmarking Apache Cassandra. I started tlp-stress back when I was working at The Last Pickle. At the time, I was spending a lot of time helping teams identify the root cause of performance issues and needed a way of benchmarking. I found cassandra-stress to be difficult to use and configure, so I ended up writing my own tool that worked in a manner that I found to be more useful. If you’re looking for a tool to assist you in benchmarking Cassandra, and you’re looking to get started quickly, this might be the right tool for you.
Back to Consulting!
Saying “it’s been a while since I wrote anything here” would be an understatement, but I’m back, with a lot to talk about in the upcoming months.
First off - if you’re not aware, I continued writing, but on The Last Pickle blog. There’s quite a few posts there, here are the most interesting ones:
- 14 Things To Do When Setting Up a New Cassandra Cluster
- Apache Cassandra Performance Tuning - Compression with Mixed Workloads
- Garbage Collection Tuning for Apache Cassandra
- Analyzing Cassandra Performance with Flame Graphs
- Cassandra Time Series Data Modeling For Massive Scale
Now the fun part - I’ve spent the last 3 years at Apple, then Netflix, neither of which gave me much time to continue my writing. As of this month, I’m officially no longer at Netflix and have started Rustyrazorblade Consulting!
Building a 100% ScyllaDB Shard-Aware Application Using Rust
Building a 100% ScyllaDB Shard-Aware Application Using Rust
I wrote a web transcript of the talk I gave with my colleagues Joseph and Yassir at [Scylla Su...
Learning Rust the hard way for a production Kafka+ScyllaDB pipeline
Learning Rust the hard way for a production Kafka+ScyllaDB pipeline
This is the web version of the talk I gave at [Scylla Summit 2022](https://www.scyllad...
On Scylla Manager Suspend & Resume feature
On Scylla Manager Suspend & Resume feature
!!! warning "Disclaimer" This blog post is neither a rant nor intended to undermine the great work that...
Renaming and reshaping Scylla tables using scylla-migrator
We have recently faced a problem where some of the first Scylla tables we created on our main production cluster were not in line any more with the evolved s...
Python scylla-driver: how we unleashed the Scylla monster's performance
At Scylla summit 2019 I had the chance to meet Israel Fruchter and we dreamed of working on adding **shard...
Scylla Summit 2019
I've had the pleasure to attend again and present at the Scylla Summit in San Francisco and the honor to be awarded the...