We Built a Better Cassandra + ScyllaDB Driver for Node.js – with Rust
Lessons learned building a Rust-backed Node.js driver for ScyllaDB: bridging JS and Rust, performance pitfalls, and benchmark results This blog post explores the story of building a new Node.js database driver as part of our Student Team Programming Project. Up ahead: troubles with bridging Rust with JavaScript, a new solution being initially a few times slower than the previous one, and a few charts! Note: We cover the progress made until June 2025 as part of the ZPP project, which is a collaboration between ScyllaDB and University of Warsaw. Since then, the ScyllaDB Driver team adopted the project (and now it’s almost production ready). Motivation The database speaks one language, but users want to speak to it in multiple languages: Rust, Go, C++, Python, JavaScript, etc. This is where a driver comes in, acting as a “translator” of sorts. All the JavaScript developers of the world currently rely on the DataStax Node.js driver. It is developed with the Cassandra database in mind, but can also be used for connecting to ScyllaDB, as they use the same protocol – CQL. This driver gets the job done, but it is not designed to take full advantage of ScyllaDB’s features (e.g., shard-per-core architecture, tablets). A solution for that is rewriting the driver and creating one that is in-house, developed and maintained by ScyllaDB developers. This is a challenging task requiring years of intensive development, with new tasks interrupting along the way. An alternative approach is writing the new driver as a wrapper around an existing one – theoretically simplifying the task (spoiler: not always) to just bridging the interfaces. This concept was proven in the making of the ScyllaDB C / C++ driver, which is an overlay over the Rust driver. We chose the ScyllaDB Rust driver as the backend of the new JavaScript driver for a few reasons. ScyllaDB’s Rust driver is developed and maintained by ScyllaDB. That means it’s always up to date with the latest database features, bug fixes, and optimizations. And since it’s written in Rust, it offers native-level performance without sacrificing memory safety. [More background on this approach] Development of such a solution skips the implementation of complicated database handling logic, but brings its own set of problems. We wanted our driver to be as similar as possible to the Node.js driver so anyone wanting to switch does not need to do much configuration. This was a restriction on one side. On the other side, we have limitations of the Rust driver interface. Driver implementations differ and the API for communicating with them can vary in some places. Some give a lot of responsibility to the user, requiring more effort but giving greater flexibility. Others do most of the work without allowing for much customization. Navigating these considerations is a recurring theme when choosing to write a driver as a wrapper over a different one. Despite the challenges during development, this approach comes with some major advantages. Once the initial integration is complete, adding new ScyllaDB features becomes much easier. It’s often just a matter of implementing a few bridging functions. All the complex internal logic is handled by the Rust driver team. That means faster development, fewer bugs, and better consistency across languages. On top of that, Rust is significantly faster than Node.js. So if we keep the overhead from the bridging layer low, the resulting driver can actually outperform existing solutions in terms of raw speed. The environment: Napi vs Napi-Rs vs Neon With the goal of creating a driver that uses ScyllaDB Rust Driver underneath, we needed to decide how we would be communicating between languages. There are two main options when it comes to communicating between JavaScript and other languages: Use a Node API (NAPI for short) – an API built directly into the NodeJS engine, or Interface the program through the V8 JavaScript engine. While we could use one of those communication methods directly, they are dedicated for C / C++, which would mean writing a lot of unsafe code. Luckily, other options exist: NAPI-RS and Neon. Those libraries handle all the unsafe code required for using the C / C++ APIs and expose (mostly safe) Rust interfaces. The first option uses NAPI exclusively under the hood, while the Neon option uses both of those interfaces. After some consideration, we decided to use NAPI-RS over Neon. Here are the things we considered when deciding which library to use: – Library approach — In NAPI-RS, the library handles the serialization of data into the expected Rust types. This lets us take full advantage of Rust’s static typing and any related optimizations. With Neon, on the other hand, we have to manually parse values into the correct types. With NAPI-RS, writing a simple function is as easy as adding a #[napi] tag: Simple a+b example And in Neon, we need to manually handle JavaScript context: A+b example in Neon – Simplicity of use — As a result of the serialization model, NAPI-RS leads to cleaner and shorter code. When we were implementing some code samples for the performance comparison, we had serious trouble implementing code in Neon just for a simple example. Based on that experience, we assumed similar issues would likely occur in the future. – Performance — We made some simple tests comparing the performance of library function calls and sending data between languages. While both options were visibly slower than pure JavaScript code, the NAPI-RS version had better performance. Since driver efficiency is a critical requirement, this was an important factor in our decision. You can read more about the benchmarks in our thesis. – Documentation — Although the documentation for both tools is far from perfect, NAPI-RS’s documentation is slightly more complete and easier to navigate. Current state and capabilities Note: This represents the state as of May 2025. More features have been introduced since then. See the project readme for a brief overview of current and planned features. The driver supports regular statements (both select and insert) and batch statements. It supports all CQL types, including encoding from almost all allowed JS types. We support prepared statements (when the driver knows the expected types based on the prepared statement), and we support unprepared statements (where users can either provide type hints, or the driver guesses expected value types). Error handling is one of the few major functions that behaves differently than the DataStax driver. Since the Rust driver throws different types of errors depending on the situation, it’s nearly impossible to map all of them reliably. To avoid losing valuable information, we pass through the original Rust errors as is. However, when errors are generated by our own logic in the wrapper, we try to keep them consistent with the old driver’s error types. In the DataStax driver, you needed to explicitly call shutdown() to close the database connection. This generated some problems: when the connection variable was dropped, the connection sometimes wouldn’t stop gracefully, even keeping the program running in some situations. We decided to switch this approach, so that the connection is automatically closed when the variable keeping the client is dropped. For now, it’s still possible to call shutdown on the client. Note: We are still discussing the right approach to handling a shutdown. As a result, the behavior described here may change in the future. Concurrent execution The driver has a dedicated endpoint for executing multiple queries concurrently. While this endpoint gives you less control over individual requests — for example, all statements must be prepared and you can’t set different options per statement — these constraints allow us to optimize performance. In fact, this approach is already more efficient than manually executing queries in parallel (around 35% faster in our internal testing), and we have additional optimization ideas planned for future implementation. Paging The Rust and DataStax drivers both have built-in support for paging, a CQL feature that allows splitting results of large queries into multiple chunks (pages). Interestingly, although the DataStax driver has multiple endpoints for paging, it doesn’t allow execution of unpaged queries. Our driver supports the paging endpoints (for now, one of those endpoints is still missing) and we also added the ability to execute unpaged queries in case someone ever needs that. With the current paging API, you have several options for retrieving paged results: Automatic iteration: You can iterate over all rows in the result set, and the driver will automatically request the next pages as needed. Manual paging: You can manually request the next page of results when you’re ready, giving you more control over the paging process. Page state transfer: You can extract the current page state and use it to fetch the next page from a different instance of the driver. This is especially useful in scenarios like stateless web servers, where requests may be handled by different server instances. Prepared statements cache Whenever executing multiple instances of the same statement, it’s recommended to use prepared statements. In ScyllaDB Rust Driver, by default, it’s the user’s responsibility to keep track of the already prepared statements to avoid preparing them multiple times (and, as a result, increasing both the network usage and execution times). In the DataStax driver, it was the driver’s responsibility to avoid preparing the same query multiple times. In the new driver, we use Rust’s Driver Caching Session for (most) of the statement caching. Optimizations One of the initial goals for the project was to have a driver that is faster than the DataStax driver. While using NAPI-RS added some overhead, we hoped the performance of the Rust driver would help us achieve this goal. With the initial implementation, we didn’t put much focus on efficient usage of the NAPI-RS layer. When we first benchmarked the new driver, it turned out to be way slower compared to both the DataStax JavaScript driver and the ScyllaDB Rust driver… Operations scylladb-javascript-driver (initial version) [s] Datastax-cassandra-driver [s] Rust-driver [s] 62500 4.08 3.53 1.04 250000 13.50 5.81 1.73 1000000 55.05 15.37 4.61 4000000 227.69 66.95 18.43 Operations scylladb-javascript-driver (initial version) [s] Datastax-cassandra-driver [s] Rust-driver [s] 62500 1.63 2.61 1.08 250000 4.09 2.89 1.52 1000000 15.74 4.90 3.45 4000000 58.96 12.72 11.64 Operations scylladb-javascript-driver (initial version) [s] Datastax-cassandra-driver [s] Rust-driver [s] 62500 1.63 2.61 1.08 250000 4.09 2.89 1.52 1000000 15.74 4.90 3.45 4000000 58.96 12.72 11.64 Operations scylladb-javascript-driver (initial version) [s] Datastax-cassandra-driver [s] Rust-driver [s] 62500 1.96 3.11 1.31 250000 4.90 4.33 1.89 1000000 16.99 10.58 4.93 4000000 65.74 31.83 17.26 Those results were a bit of a surprise, as we didn’t fully anticipate how much overhead NAPI-RS would introduce. It turns out that using JavaScript Objects introduced way higher overhead compared to other built-in types, or Buffers. You can see on the following flame graph how much time was spent executing NAPI functions (yellow-orange highlight), which are related to sending objects between languages. Creating objects with NAPI-RS is as simple as adding the
#[napi] tag to the struct we want to expose to the
NodeJS part of the code. This approach also allows us to create
methods on those objects. Unfortunately, given its overhead, we
needed to switch the approach – especially in the most used parts
of the driver, like parsing parameters, results, or other parts of
executing queries. We can create a napi object like this:
Which is converted to the following JavaScript class: We
can use this struct between JavaScript and Rust. When accepting
values as arguments to Rust functions exposed in NAPI-RS, we can
either accept values of the types that implement the
FromNapiValue trait, or accept references to values of
types that are exposed to NAPI (these implement the default
FromNapiReference trait). We can do it like this:
Then, when we call the following Rust function we
can just pass a number in the JavaScript code.
FromNapiValue is implemented for built-in types like
numbers or strings, and the
FromNapiReference trait is created automatically
when using the #[napi] tag on a Rust struct. Compared
to that, we need to manually implement
FromNapiValue for custom structs. However, this
approach allows us to receive those objects in functions exposed to
NodeJS, without the need for creating Objects – and thus
significantly improves performance. We used this mostly to improve
the performance of passing query parameters to the Rust side of the
driver. When it comes to returning values from Rust code, a type
must have a ToNapiValue trait implemented. Similarly,
this trait is already implemented for built-in types, and is auto
generated with macros when adding the #[napi] tag to
the object. And this auto generated implementation was causing most
of our performance problems. Luckily, we can also implement our own
ToNapiValue trait. If we return a raw value and create
an object directly in the JavaScript part of the code, we can avoid
almost all of the negative performance impacts that come from the
default implementation of ToNapiValue. We can do it
like this: 
ToNapiValue implementation, that
was handling all the logic for the UUID. When we changed the
approach to returning just a raw buffer representing the UUID and
handling all the logic on the JavaScript side, we shaved off about
20% of the CPU time we were using in the select benchmarks at that
point in time. Note: Since completing the initial project, we’ve
introduced additional changes to how serialization and
deserialization works. This means the current state may be
different from what we describe here. A new round of benchmarking
is in progress; stay tuned for those results. Benchmarks In the
previous section, we showed you some early benchmarks. Let’s talk a
bit more about how we tested and what we tested. All benchmarks
presented here were run on a single machine – the database was run
in a Docker container and the driver benchmarks were run without
any virtualization or containerization. The machine was running on
AMD Ryzen™ 7 PRO 7840U with 32GB RAM, with the database itself
limited to 8GB of RAM in total. We tested the driver both with
ScyllaDB and Cassandra (latest stable versions as of the time of
testing – May 2025). Both of those databases were run in a three
node configuration, with 2 shards per node in the case of ScyllaDB.
With this information on the benchmarks, let’s see the effect all
the optimizations we added had on the driver performance when
tested with ScyllaDB:
Operations Scylladb-javascript-driver [s] Datastax-cassandra-driver
[s] Rust-driver [s] scylladb-javascript-driver (initial version)
[s] 62500 1.89 3.45 0.99 4.08 250000 4.15 5.66 1.73 13.50 1000000
13.65 15.86 4.41 55.05 4000000 55.85 56.73 18.42 227.69
Operations Scylladb-javascript-driver [s] Datastax-cassandra-driver
[s] Rust-driver [s] scylladb-javascript-driver (initial version)
[s] 62500 2.83 2.48 1.04 1.63 250000 1.91 2.91 1.56 4.09 1000000
4.58 4.69 3.42 15.74 4000000 16.05 14.27 11.92 58.96
Operations Scylladb-javascript-driver [s] Datastax-cassandra-driver
[s] Rust-driver [s] scylladb-javascript-driver (initial version)
[s] 62500 1.50 3.04 1.33 1.96 250000 2.93 4.52 1.94 4.90 1000000
8.79 11.11 5.08 16.99 4000000 32.99 36.62 17.90 65.74
Operations Scylladb-javascript-driver [s] Datastax-cassandra-driver
[s] Rust-driver [s] scylladb-javascript-driver (initial version)
[s] 62500 1.42 3.09 1.25 1.45 250000 2.94 3.81 2.43 3.43 1000000
9.19 8.98 7.21 10.82 4000000 33.51 28.97 25.81 40.74 And here are
the same benchmarks, without the initial driver version.
Here are the results of running the benchmark on Cassandra.
Operations Scylladb-javascript-driver [s]
Datastax-cassandra-driver [s] Rust-driver [s] 62500 2.48 14.50 1.25
250000 5.82 19.93 2.00 1000000 19.77 19.54 5.16
Operations Scylladb-javascript-driver [s]
Datastax-cassandra-driver [s] Rust-driver [s] 62500 1.60 2.99 1.48
250000 3.06 4.46 2.42 1000000 9.02 9.03 6.53
Operations Scylladb-javascript-driver [s] Datastax-cassandra-driver
[s] Rust-driver [s] 62500 2.32 4.03 2.11 250000 5.45 6.53 4.01
1000000 18.77 16.20 13.21
Operations Scylladb-javascript-driver [s] Datastax-cassandra-driver
[s] Rust-driver [s] 62500 1.86 4.15 1.57 250000 4.24 5.41 3.36
1000000 13.11 14.11 10.54 The test results across both ScyllaDB and
Cassandra show that the new driver has slightly better performance
on the insert benchmarks. For select benchmarks, it starts ahead
and the performance advantage decreases with time. Despite a series
of optimizations, the majority of the CPU time still comes from
NAPI communication and thread synchronization (according to
internal flamegraph testing). There is still some room for
improvement, which we’re going to explore. Since running those
benchmarks, we introduced changes that improve the performance of
the driver. With those improvements performance of select
benchmarks is much closer to the speed of the DataStax driver.
Again…please stay tuned for another blog post with updated results.
Shards and tablets Since the DataStax driver lacked tablet and
shard support, we were curious if our new shard-aware and
tablet-aware drivers provided a measurable performance gain with
shards and tablets.
Operations ScyllaDB JS Driver [s] DataStax Driver [s] Rust Driver
[s] Shard-Aware No Shards Shard-Aware No Shards Shard-Aware No
Shards 62,500 1.89 2.61 3.45 3.51 0.99 1.20 250,000 4.15 7.61 5.66
6.14 1.73 2.30 1,000,000 13.65 30.36 15.86 16.62 4.41 8.33
4,000,000 55.85 134.90 56.73 77.68 18.42 42.64
Operations ScyllaDB JS Driver [s] DataStax Driver [s] Rust Driver
[s] Shard-Aware No Shards Shard-Aware No Shards Shard-Aware No
Shards 62,500 1.50 1.52 3.04 3.63 1.33 1.33 250,000 2.93 3.29 4.52
5.09 1.94 2.02 1,000,000 8.79 10.29 11.11 11.13 5.08 5.71 4,000,000
32.99 38.53 36.62 39.28 17.90 20.67 In insert benchmarks, there are
noticeable changes across all drivers when having more than one
shard. The Rust driver improved by around 36%, the new driver
improved by around 46%, and the DataStax driver improved by only
around 10% when compared to the single sharded version. While
sharding provides some performance benefits for the DataStax
driver, which is not shard aware, the new driver benefits
significantly more — achieving performance improvements comparable
to the Rust driver. This shows that it’s not only introducing more
shards that provide an improvement in this case; a major part of
the performance improvement is indeed shard-awareness.
Operations ScyllaDB JS Driver [s] DataStax Driver [s] Rust Driver
[s] No Tablets Standard No Tablets Standard No Tablets Standard
62,500 1.76 1.89 3.67 3.45 1.06 0.99 250,000 3.91 4.15 5.65 5.66
1.59 1.73 1,000,000 12.81 13.65 13.54 15.86 3.74 4.41
Operations ScyllaDB JS Driver [s] DataStax Driver [s] Rust Driver
[s] No Tablets Standard No Tablets Standard No Tablets Standard
62,500 1.46 1.50 2.92 3.04 1.33 1.33 250,000 2.76 2.93 4.03 4.52
1.94 1.94 1,000,000 8.36 8.79 7.68 11.11 4.84 5.08 When it comes to
tablets, the new driver and the Rust driver see only minimal
changes to the performance, while the performance of the DataStax
driver drops significantly. This behavior is expected. The DataStax
driver is not aware of the tablets. As a result, it is unable to
communicate directly with the node that will store the data – and
that increases the time spent waiting on network communication.
Interesting things happen, however, when we look at the network
traffic: WHAT TOTAL CQL TCP Total Size New driver 3 node all
412,764 112,318 300,446 ∼ 43.7 MB New driver 3 node | driver ↔
database 409,678 112,318 297,360 – New driver 3 node | node ↔ node
3,086 0 3,086 – DataStax driver 3 node all 268,037 45,052 222,985 ∼
81.2 MB DataStax driver 3 node | driver ↔ database 90,978 45,052
45,926 – DataStax driver 3 node | node ↔ node 177,059 0 177,059 –
This table shows the number of packets sent during the concurrent
insert benchmark on three-node ScyllaDB with 2 shards per node.
Those results were obtained with RF = 1. While running the database
with such a replication factor is not production-suitable, we
chose it to better visualize the results. When looking at those
numbers, we can draw the following conclusions: The new driver has
a different coalescing mechanism. It has a shorter wait time, which
means it sends more messages to the database and achieves
lower latencies. The new driver knows which node(s) will
store the data. This reduces internal traffic between database
nodes and lets the database serve more traffic with the same
resources. Future plans The goal of this project was to create a
working prototype, which we managed to successfully achieve. It’s
available at https://github.com/scylladb/nodejs-rs-driver,
but it’s considered experimental at this point. Expect it to change
considerably, with ongoing work and refactors. Some of the features
that were present in DataStax driver, and are expected for the
driver to be considered deployment-ready, are not yet implemented.
The Drivers team is actively working to add those features. If
you’re interested in this project and would like to contribute,
here’s the project’s GitHub
repository. 
Exploring the key features of Cassandra® 5.0
Apache Cassandra has become one of the most broadly adopted distributed databases for large-scale, highly available applications since its launch as an open source project in 2008. The 5.0 release in September 2024 represents the most substantial advancement to the project since 4.0 released in July 2021. Multiple customers (and our own internal Cassandra use case) have now been happily running on Cassandra 5 for up to 12 months so we thought the time was right to explore the key features they are leveraging to power their modern applications.
An overview of new features in Apache Cassandra 5.0Apache Cassandra 5.0 introduces core capabilities aimed at AI-driven systems, low-latency analytical workloads, and environments that blend operational and analytical processing.
Highlights include:
- The new vector data type and an Approximate Nearest Neighbor (ANN) index based on Hierarchical Navigable Small World (HNSW), which is integrated into the Storage-Attached Index (SAI) architecture
- Trie-based memtables and the Big Trie-Index (BTI) SSTable format, delivering better memory efficiency and more consistent write performance
- The Unified Compaction Strategy, a tunable density-based approach that can align with leveled or tiered compaction patterns.
Additional enhancements include expanded mathematical CQL functions, dynamic data masking, and experimental support for Java 17.
At NetApp, Apache Cassandra 5.0 is fully supported, and we are actively assisting customers as they transition from 4.x.
A deeper look at Cassandra 5.0’s key features Storage–Attached Indexes (SAI)Storage–Attached Indexes bring a modern, storage-integrated approach to secondary indexing in Apache Cassandra, resolving many of the scalability and maintenance challenges associated with earlier index implementations. Legacy Secondary Indexes (2i) and SASI remain available, but SAI offers a more robust and predictable indexing model for a broad range of production workloads.
SAI operates per-SSTable, allowing queries to be indexed locally versus the cluster-wide coordination required of other strategies. This model supports diverse CQL data types, enables efficient numeric and text range filters, and provides more consistent performance characteristics than 2i or SASI. The same storage-attached foundation is also used for Cassandra 5’s vector indexing mechanism, allowing ANN search to operate within the same storage and query framework.
SAI supports combining filters across multiple indexed columns and works seamlessly with token-aware routing to reduce unnecessary coordinator work. Public evaluations and community testing have shown faster index builds, more predictable read paths, and improved disk utilization compared with previous index formats.
Operationally, SAI functions as part of the storage engine itself: indexes are defined using standard CQL statements and are maintained automatically during flush and compaction, with no cluster-wide rebuilds required. This provides more flexible query options and can simplify application designs that previously relied on manual denormalization or external indexing systems.
Native Vector Search capabilitiesApache Cassandra 5.0 introduces native support for high-dimensional vector embeddings through the new vector data type. Embeddings represent semantic information in numerical form, enabling similarity search to be performed directly within the database. The vector type is integrated with the database’s storage-attached index architecture, which uses HNSW graphs to efficiently support ANN search across cosine, Euclidean, and dot-product similarity metrics.
With vector search implemented at the storage layer, applications involving semantic matching, content discovery, and retrieval-oriented workflows while maintaining the system’s established scalability and fault-tolerance characteristics are supported.
After upgrading to 5.0, existing schemas can add vector columns and store embeddings through standard write operations. For example:
UPDATE products SET embedding = [0.1, 0.2, 0.3, 0.4, 0.5] WHERE id = <id>;
To create a new table with a vector type column:
CREATE TABLE items ( product_id UUID PRIMARY KEY, embedding VECTOR<FLOAT, 768> // 768 denotes dimensionality );
Because vector indexes are attached to SSTables, they participate automatically in the compaction and repair processes and do not require an external indexing system. ANN queries can be combined with regular CQL filters, allowing similarity searches and metadata conditions to be evaluated within a unified distributed query workflow. This brings vector retrieval into Apache Cassandra’s native consistency, replication, and storage model.
Unified Compaction Strategy (UCS)Unified Compaction Strategy in Apache Cassandra 5 included a density-aware approach to organizing SSTables that blends the strengths of Leveled Compaction Strategy (LCS) and Size Tiered Compaction Strategy (STCS). UCS aims to provide the predictable read amplification associated with LCS and the write efficiency of STCS, without many of the workload-specific drawbacks that previously made compaction selection difficult. Choosing an unsuitable compaction strategy in earlier releases could lead to operational complexity and long-term performance issues, which UCS is designed to mitigate.
UCS exposes a set of tunable parameters like density thresholds and per-level scaling that let operators adjust compaction behavior toward read-heavy, write-heavy, or time-series patterns. This flexibility also helps smooth the transition from existing strategies, as UCS can adopt and improve the current SSTable layout without requiring a full rewrite in most cases. The introduction of compaction shards further increases parallelism and reduces the impact of large compactions on cluster performance.
Although LCS and STCS remain available (and while STCS remains the default strategy in 5.0, UCS is the default strategy on newly deployed NetApp Instaclustr’s managed Apache Cassandra 5 clusters), UCS supports a broader range of workloads, reduces the operational burden of compaction tuning, and aligns well with other storage engine improvements in Apache Cassandra 5 such as trie-based SSTables and Storage-Attached Indexes.
Trie Memtables and Trie-Indexed SSTablesTrie Memtables and Trie-indexed SSTables (Big Trie-Index, BTI) are significant storage engine enhancements released in Apache Cassandra 5. They are designed to reduce memory overhead, improve lookup performance, and increase flush efficiency. A trie data structure stores keys by shared prefixes instead of repeatedly storing full keys, which lowers object count and improves CPU cache locality compared with the legacy skip-list memtable structure. These benefits are particularly visible in high-ingestion, IoT, and time-series workloads.
Skip-list memtables store full keys for every entry, which can lead to large heap usage and increased garbage collection activity under heavy write loads. Trie Memtables substantially reduce this overhead by compacting key storage and avoiding pointer-heavy layouts. On disk, the BTI SSTable format replaces the older BIG index with a trie-based partition index that removes redundant key material and reduces the number of key comparisons needed during partition lookups.
Using Trie memtables requires enabling both the trie-based memtable implementation and the BTI SSTable format. Existing BIG SSTables are converted to BTI through normal compaction or by rebuilding data. On NetApp Instaclustr’s managed Apache Cassandra clusters Trie Memtables and BTI are enabled by default, but when upgrading major versions to 5.0, data must be converted from BIG to BTI first to utilize Trie structures.
Other new features Mathematical CQL functionsApache Cassandra 5.0 added a rich set of math functions allowing developers to perform computations directly within queries. This reduces data transfer overhead and reduces client-side post-processing, among many other benefits. From fundamental functions like ABS(), ROUND(), or SQRT() to more complex operations like SIN(), COS(), TAN(), these math functions are extensible to a multitude of domains from financial data, scientific measurements or spatial data.
Dynamic Data MaskingDynamic Data Masking (DDM) is a new feature to obscure sensitive
column-level data at query time or permanently attach the
functionality to a column so that the data always returns
obfuscated. Stored data values are not altered in this process, and
administrators can control access through role-based access control
(RBAC) to ensure only those with access can see the data while also
tuning the visibility of the obscured data. This feature helps with
adherence to data privacy regulations such as GDPR, HIPAA, and PCI
DSS without needing external redaction systems.
Apache Cassandra 5.0 packs a punch with game changing features that meet the needs of modern workloads and applications. Features like vector search capabilities and Storage Attached Indexes stand out as they will inevitably shape how data can be leveraged within the same database while maintaining speed, scale, and resilience.
When you deploy a managed cluster on NetApp Instaclustr’s Managed Platform, you get the benefits of all these amazing features without worrying about configuration and maintenance.
Ready to experience the power of Apache Cassandra 5.0 for yourself? Try it free for 30 days today!
The post Exploring the key features of Cassandra® 5.0 appeared first on Instaclustr.
Instaclustr product update: December 2025
Instaclustr product update: December 2025Here’s a roundup of the latest features and updates that we’ve recently released.
If you have any particular feature requests or enhancement ideas that you would like to see, please get in touch with us.
Major announcements OpenSearch®AI Search for OpenSearch®: Unlocking next-generation search
AI Search for OpenSearch, which is now available in Public Preview on the NetApp Instaclustr Managed Platform, is designed to bring semantic, hybrid, and multimodal search capabilities to OpenSearch deployments—turning them into an end-to-end AI-powered search solution within minutes. With built-in ML models, vector indexing, and streamlined ingestion pipelines, next-generation search can be enabled in minutes without adding operational complexity. This feature powers smarter, more relevant discovery experiences backed by AI—securely deployed across any cloud or on-premises environment.
ClickHouse®
FSx for NetApp ONTAP and Managed ClickHouse® integration is now
available
We’re excited to announce that NetApp has introduced seamless
integration between Amazon FSx for NetApp ONTAP and Instaclustr
Managed ClickHouse, to enable customers to build a truly hybrid
lakehouse architecture on AWS. This integration is designed to
deliver lightning-fast analytics without the need for complex data
movement, while leveraging FSx for ONTAP’s unified file and object
storage, tiered performance, and cost optimization. Customers can
now run zero-copy lakehouse analytics with ClickHouse directly on
FSx for ONTAP data—to simplify operations, accelerate
time-to-insight, and reduce total cost of ownership.
Instaclustr for PostgreSQL® on Amazon FSx for ONTAP: A new
era
We’re excited to announce the public preview of Instaclustr Managed
PostgreSQL integrated with Amazon FSx for NetApp ONTAP—combining
enterprise-grade storage with world-class open source database
management. This integration is designed to deliver higher IOPS,
lower latency, and advanced data management without increasing
instance size or adding costly hardware. Customers can now run
PostgreSQL clusters backed by FSx for ONTAP storage, leveraging
on-disk compression for cost savings and paving the way for
ONTAP-powered features, such as instant snapshot backups, instant
restores, and fast forking. These ONTAP-enabled features are
planned to unlock huge operational benefits and will be launched
with our GA release.
- Released Apache Cassandra 5.0.5 into general availability on the NetApp Instaclustr Managed Platform.
- Transitioned Apache Cassandra v4.1.8 to CLOSED lifecycle state; scheduled to reach End of Life (EOL) on December 20, 2025.
- Kafka on Azure now supports v5 generation nodes, available in General Availability.
- Instaclustr Managed Apache ZooKeeper has moved from General Availability to closed status.
- Kafka and Kafka Connect 3.1.2 and 3.5.1 are retired; 3.6.2, 3.7.1, 3.8.1 are in legacy support. Next set of lifecycle state changes for Kafka and Kafka Connect in end March 2026 will see all supported versions 3.8.1 and below marked End of Life.
- Karapace Rest Proxy and Schema Registry 3.15.0 are closed. Customers are advised to move to version 5.x.
- Kafka Rest Proxy 5.0.0 and Kafka Schema Registry 5.0.0, 5.0.4 have been moved to end of life. Affected customers have been contacted by Support to schedule a migration to a supported version as soon as possible.
- ClickHouse 25.3.6 has been added to our managed platform in General Availability.
- Kafka Table Engine integration with ClickHouse has added support to enable real-time data ingestion, streamline streaming analytics, and accelerate insights.
- New ClickHouse node sizes, powered by AWS m7g, r7i, and r7g instances, are now in Limited Availability for cluster creation.
- Cadence is now available to be provisioned with Cassandra 5.x, designed to deliver improved performance, enhanced scalability, and stronger security for mission-critical workflows.
- OpenSearch 2.19.3 and 3.2.0 have been released to General Availability.
- PostgreSQL AWS PrivateLink support has been added, enabling connectivity between VPCs using AWS PrivateLink.
- PostgreSQL version 18.0 has now been released to General Availability, alongside PostgreSQL version 16.10, 17.6.
- Added new PostgreSQL metrics for connect states and wait event types.
- PostgreSQL Load Balancer add-on is now available, providing a unified endpoint for cluster access, simplifying failover handling, and ensuring node health through regular checks.
- We’re working on enabling multi-datacenter (multi-DC) cluster provisioning via API and console, designed to make it easier to deploy clusters across regions with secure networking and reduced manual steps.
- We’re working on adding Kafka Tiered Storage for clusters running in GCP— designed to bring affordable, scalable retention, and instant access to historical data, to ensure flexibility and performance across clouds for enterprise Kafka users.
- We’re planning to extend our Managed ClickHouse to allow it to work with on-prem deployments.
- Following the success of our public preview, we’re preparing to launch PostgreSQL integrated with FSx for NetApp ONTAP (FSxN) into General Availability. This enhancement is designed to combine enterprise-grade PostgreSQL with FSxN’s scalable, cost-efficient storage, enabling customers to optimize infrastructure costs while improving performance and flexibility.
- As part of our ongoing advancements in AI for OpenSearch, we are planning to enable adding GPU nodes into OpenSearch clusters, aiming to enhance the performance and efficiency of machine learning and AI workloads.
- Self-service Tags Management feature—allowing users to add, edit, or delete tags for their clusters directly through the Instaclustr console, APIs, or Terraform provider for RIYOA deployments.
- Cadence Workflow, the open source orchestration engine created by Uber, has officially joined the Cloud Native Computing Foundation (CNCF) as a Sandbox project. This milestone ensures transparent governance, community-driven innovation, and a sustainable future for one of the most trusted workflow technologies in modern microservices and agentic AI architectures. Uber donates Cadence Workflow to CNCF: The next big leap for the open source project—read the full story and discover what’s next for Cadence.
- Upgrading ClickHouse® isn’t just about new features—it’s essential for security, performance, and long-term stability. In ClickHouse upgrade: Why staying updated matters, you’ll learn why skipping upgrades can lead to technical debt, missed optimizations, and security risks. Then, explore A guide to ClickHouse® upgrades and best practices for practical strategies, including when to choose LTS releases for mission-critical workloads and when stable releases make sense for fast-moving environments.
- Our latest blog, AI Search for OpenSearch®: Unlocking next-generation search, explains how this new solution enables smarter discovery experiences using built-in ML models, vector embeddings, and advanced search techniques—all fully managed on the NetApp Instaclustr Platform. Ready to explore the future of search? Read the full article and see how AI can transform your OpenSearch deployments.
If you have any questions or need further assistance with these enhancements to the Instaclustr Managed Platform, please contact us.
SAFE HARBOR STATEMENT: Any unreleased services or features referenced in this blog are not currently available and may not be made generally available on time or at all, as may be determined in NetApp’s sole discretion. Any such referenced services or features do not represent promises to deliver, commitments, or obligations of NetApp and may not be incorporated into any contract. Customers should make their purchase decisions based upon services and features that are currently generally available.
The post Instaclustr product update: December 2025 appeared first on Instaclustr.
Freezing streaming data into Apache Iceberg™—Part 1: Using Apache Kafka®Connect Iceberg Sink Connector
IntroductionEver since the first distributed system—i.e. 2 or more computers networked together (in 1969)—there has been the problem of distributed data consistency: How can you ensure that data from one computer is available and consistent with the second (and more) computers? This problem can be uni-directional (one computer is considered the source of truth, others are just copies), or bi-directional (data must be synchronized in both directions across multiple computers).
Some approaches to this problem I’ve come across in the last 8 years include Kafka Connect (for elegantly solving the heterogeneous many-to-many integration problem by streaming data from source systems to Kafka and from Kafka to sink systems, some earlier blogs on Apache Camel Kafka Connectors and a blog series on zero-code data pipelines), MirrorMaker2 (MM2, for replicating Kafka clusters, a 2 part blog series), and Debezium (Change Data Capture/CDC, for capturing changes from databases as streams and making them available in downstream systems, e.g. for Apache Cassandra and PostgreSQL)—MM2 and Debezium are actually both built on Kafka Connect.
Recently, some “sink” systems have been taking over responsibility for streaming data from Kafka into themselves, e.g. OpenSearch pull-based ingestion (c.f. OpenSearch Sink Connector), and the ClickHouse Kafka Table Engine (c.f. ClickHouse Sink Connector). These “pull-based” approaches are potentially easier to configure and don’t require running a separate Kafka Connect cluster and sink connectors, but some downsides may be that they are not as reliable or independently scalable, and you will need to carefully monitor and scale them to ensure they perform adequately.
And then there’s “zero-copy” approaches—these rely on the well-known computer science trick of sharing a single copy of data using references (or pointers), rather than duplicating the data. This idea has been around for almost as long as computers, and is still widely applicable, as we’ll see in part 2 of the blog.
The distributed data use case we’re going to explore in this 2-part blog series is streaming Apache Kafka data into Apache Iceberg, or “Freezing streaming Apache Kafka data into an (Apache) Iceberg”! In part 1 we’ll introduce Apache Iceberg and look at the first approach for “freezing” streaming data using the Kafka Connect Iceberg Sink Connector.
What is Apache Iceberg?Apache Iceberg is an open source specification open table format optimized for column-oriented workloads, supporting huge analytic datasets. It supports multiple different concurrent engines that can insert and query table data using SQL—and Iceberg is organized like, well, an iceberg!
The tip of the Iceberg is the Catalog. An Iceberg Catalog acts as a central metadata repository, tracking the current state of Iceberg tables, including their names, schemas, and metadata file locations. It serves as the “single source of truth” for a data Lakehouse, enabling query engines to find the correct metadata file for a table to ensure consistent and atomic read/write operations.
Just under the water, the next layer is the metadata layer. The Iceberg metadata layer tracks the structure and content of data tables in a data lake, enabling features like efficient query planning, versioning, and schema evolution. It does this by maintaining a layered structure of metadata files, manifest lists, and manifest files that store information about table schemas, partitions, and data files, allowing query engines to prune unnecessary files and perform operations atomically.
The data layer is at the bottom. The Iceberg data layer is the storage component where the actual data files are stored. It supports different storage backends, including cloud-based object storage like Amazon S3 or Google Cloud Storage, or HDFS. It uses file formats like Parquet or Avro. Its main purpose is to work in conjunction with Iceberg’s metadata layer to manage table snapshots and provide a more reliable and performant table format for data lakes, bringing data warehouse features to large datasets.
As shown in the above diagram, Iceberg supports multiple different engines, including Apache Spark and ClickHouse. Engines provide the “database” features you would expect, including:
- Data Management
- ACID Transactions
- Query Planning and Optimization
- Schema Evolution
- And more!
I’ve recently been reading an excellent book on Apache Iceberg (“Apache Iceberg: The Definitive Guide”), which explains the philosophy, architecture and design, including operation, of Iceberg. For example, it says that it’s best practice to treat data lake storage as immutable—data should only be added to a Data Lake, not deleted. So, in theory at least, writing infinite, immutable Kafka streams to Iceberg should be straightforward!
But because it’s a complex distributed system (which looks like a database from above water but is really a bunch of files below water!), there is some operational complexity. For example, it handles change and consistency by creating new snapshots for every modification, enabling time travel, isolating readers from writes, and supporting optimistic concurrency control for multiple writers. But you need to manage snapshots (e.g. expiring old snapshots). And chapter 4 (performance optimisation) explains that you may need to worry about compaction (reducing too many small files), partitioning approaches (which can impact read performance), and handling row-level updates. The first two issues may be relevant for Kafka, but probably not the last one. So, it looks like it’s good fit for the streaming Kafka use cases, but we may need to watch out for Iceberg management issues.
“Freezing” streaming data with the Kafka Iceberg Sink ConnectorBut Apache Iceberg is “frozen”—what’s the connection to fast-moving streaming data? You certainly don’t want to collide with an iceberg from your speedy streaming “ship”—but you may want to freeze your streaming data for long-term analytical queries in the future. How can you do that without sinking? Actually, a “sink” is the first answer: A Kafka Connect Iceberg Sink Connector is the most common way of “freezing” your streaming data in Iceberg!
Kafka Connect is the standard framework provided by Apache Kafka to move data from multiple heterogeneous source systems to multiple heterogeneous sink systems, using:
- A Kafka cluster
- A Kafka Connect cluster (running connectors)
- Kafka Connect source connectors
- Kafka topics and
- Kafka Connect Sink Connectors
That is, a highly decoupled approach. It provides real-time data movement with high scalability, reliability, error handling and simple transformations.
Here’s the Kafka Connect Iceberg Sink Connector official documentation.
It appears to be reasonably complicated to configure this sink connector; you will need to know something about Iceberg. For example, what is a “control topic”? It’s apparently used to coordinate commits for exactly-once semantics (EOS).
The connector supports fan-out (writing to multiple Iceberg tables from one topic), fan-in (writing to one Iceberg table from multiple topics), static and dynamic routing, and filtering.
In common with many technologies that you may want to use as Kafka Connect sinks, they may not all have good support for Kafka metadata. The KafkaMetadata Transform (which injects topic, partition, offset and timestamp properties) is only experimental at present.
How are Iceberg tables created with the correct metadata? If you have JSON record values, then schemas are inferred by default (but may not be correct or optimal). Alternatively, explicit schemas can be included in-line or referenced from a Kafka Schema Registry (e.g. Karapace), and, as an added bonus, schema evolution is supported. Also note that Iceberg tables may have to be manually created prior to use if your Catalog doesn’t support table auto-creation.
From what I understood about Iceberg, to use it (e.g. for writes), you need support from an engine (e.g. to add raw data to the Iceberg warehouse, create the metadata files, and update the catalog). How does this work for Kafka Connect? From this blog I discovered that the Kafka Connect Iceberg Sink connector is functioning as an Iceberg engine for writes, so there really is an engine, but it’s built into the connector.
As is the case with all Kafka Connect Sink Connectors, records are available immediately they are written to Kafka topics by Kafka producers and Kafka Connect Source Connectors, i.e. records in active segments can be copied immediately to sink systems. But is the Iceberg Sink Connector real-time? Not really! The default time to write to Iceberg is every 5 minutes (iceberg.control.commit.interval-ms) to prevent multiplication of small files—something that Iceberg(s) doesn’t/don’t like (“melting”?). In practice, it’s because every data file must be tracked in the metadata layer, which impacts performance in many ways—proliferation of small files is typically addressed by optimization and compaction (e.g. Apache Spark supports Iceberg management, including these operations).
So, unlike most Kafka Connect sink connectors, which write as quickly as possible, there will be lag before records appear in Iceberg tables (“time to freeze” perhaps)!
The systems are separate (Kafka and Iceberg are independent), records are copied to Iceberg, and that’s it! This is a clean separation of concerns and ownership. Kafka owns the source data (with Kafka controlling data lifecycles, including record expiry), Kafka Connect Iceberg Sink Connector performs the reading from Kafka and writing to Iceberg, and is independently scalable to Kafka. Kafka doesn’t handle any of the Iceberg management. Once the data has landed in Iceberg, Kafka has no further visibility or interest in it. And the pipeline is purely one way, write only – reads or deletes are not supported.
Here’s a summary of this approach to freezing streams:
- Kafka Connect Iceberg Sink Connector shares all the benefits of the Kafka Connect framework, including scalability, reliability, error handling, routing, and transformations.
- At least, JSON values are required, ideally full schemas and referenced in Karapace—but not all schemas are guaranteed to work.
- Kafka Connect doesn’t “manage” Iceberg (e.g. automatically aggregate small files, remove snapshots, etc.)
- You may have to tune the commit interval – 5 minutes is the default.
- But it does have a built-in engine that supports writing to Iceberg.
- You may need to use an external tool (e.g. Apache Spark) for Iceberg management procedures.
- It’s write-only to Iceberg. Reads or deletes are not supported
But what’s the best thing about the Kafka Connect Iceberg Sink Connector? It’s available now (as part of the Apache Iceberg build) and works on the NetApp Instaclustr Kafka Connect platform as a “bring your own connector” (instructions here).
In part 2, we’ll look at Kafka Tiered Storage and Iceberg Topics!
The post Freezing streaming data into Apache Iceberg™—Part 1: Using Apache Kafka®Connect Iceberg Sink Connector appeared first on Instaclustr.
Stay ahead with Apache Cassandra®: 2025 CEP highlights
Apache Cassandra® committers are working hard, building new features to help you more seamlessly ease operational challenges of a distributed database. Let’s dive into some recently approved CEPs and explain how these upcoming features will improve your workflow and efficiency.
What is a CEP?CEP stands for Cassandra Enhancement Proposal. They are the process for outlining, discussing, and gathering endorsements for a new feature in Cassandra. They’re more than a feature request; those who put forth a CEP have intent to build the feature, and the proposal encourages a high amount of collaboration with the Cassandra contributors.
The CEPs discussed here were recently approved for implementation or have had significant progress in their implementation. As with all open-source development, inclusion in a future release is contingent upon successful implementation, community consensus, testing, and approval by project committers.
CEP-42: Constraints frameworkWith collaboration from NetApp Instaclustr, CEP-42, and subsequent iterations, delivers schema level constraints giving Cassandra users and operators more control over their data. Adding constraints on the schema level means that data can be validated at write time and send the appropriate error when data is invalid.
Constraints are defined in-line or as a separate definition. The inline style allows for only one constraint while a definition allows users to define multiple constraints with different expressions.
The scope of this CEP-42 initially supported a few constraints that covered the majority of cases, but in follow up efforts the expanded list of support includes scalar (>, <, >=, <=), LENGTH(), OCTET_LENGTH(), NOT NULL, JSON(), REGEX(). A user is also able to define their own constraints if they implement it and put them on Cassandra’s class path.
A simple example of an in-line constraint:
CREATE TABLE users (
username text PRIMARY KEY,
age int CHECK age >= 0 and age < 120
);
Constraints are not supported for UDTs (User-Defined Types) nor collections (except for using NOT NULL for frozen collections).
Enabling constraints closer to the data is a subtle but mighty way for operators to ensure that data goes into the database correctly. By defining rules just once, application code is simplified, more robust, and prevents validation from being bypassed. Those who have worked with MySQL, Postgres, or MongoDB will enjoy this addition to Cassandra.
CEP-51: Support “Include” Semantics for cassandra.yamlThe cassandra.yaml file holds important settings for storage,
memory, replication, compaction, and more. It’s no surprise that
the average size of the file around 1,000 lines (though, yes—most
are comments). CEP-51 enables splitting the cassandra.yaml
configuration into multiple files using includes
semantics. From the outside, this feels like a small change, but
the implications are huge if a user chooses to opt-in.
In general, the size of the configuration file makes it difficult to manage and coordinate changes. It’s often the case that multiple teams manage various aspects of the single file. In addition, cassandra.yaml permissions are readable for those with access to this file, meaning private information like credentials are comingled with all other settings. There is risk from an operational and security standpoint.
Enabling the new semantics and therefore modularity for the configuration file eases management, deployment, complexity around environment-specific settings, and security in one shot. The configuration file follows the principle of least privilege once the cassandra.yaml is broken up into smaller, well-defined files; sensitive configuration settings are separated out from general settings with fine-grained access for the individual files. With the feature enabled, different development teams are better equipped to deploy safely and independently.
If you’ve deployed your Cassandra cluster on the NetApp Instaclustr platform, the cassandra.yaml file is already configured and managed for you. We pride ourselves on making it easy for you to get up and running fast.
CEP-52: Schema annotations for Apache CassandraWith extensive review by the NetApp Instaclustr team and Stefan Miklosovic, CEP-52 introduces schema annotations in CQL allowing in-line comments and labels of schema elements such as keyspaces, tables, columns, and User Defined Types (UDT). Users can easily define and alter comments and labels on these elements. They can be copied over when desired using CREATE TABLE LIKE syntax. Comments are stored as plain text while labels are stored as structured metadata.
Comments and labels serve different annotation purposes: Comments document what a schema object is for, whereas labels describe how sensitive or controlled it is meant to be. For example, labels can be used to identify columns as “PII” or “confidential”, while the comment on that column explains usage, e.g. “Last login timestamp.”
Users can query these annotations. CEP-52 defines two new read-only tables (system_views.schema_comments and system_views.schema_security_labels) to store comments and security labels so objects with comments can be returned as a list or a user/machine process can query for specific labels, beneficial for auditing and classification. Note that adding security labels are descriptive metadata and do not enforce access control to the data.
CEP-53: Cassandra rolling restarts via SidecarSidecar is an auxiliary component in the Cassandra ecosystem that exposes cluster management and streaming capabilities through APIs. Introducing rolling restarts through Sidecar, this feature is designed to provide operators with more efficient and safer restarts without cluster-wide downtime. More specifically, operators can monitor, pause, resume, and abort restarts all through an API with configurable options if restarts fail.
Rolling restarts brings operators a step closer to cluster-wide operations and lifecycle management via Sidecar. Operators will be able to configure the number of nodes to restart concurrently with minimal risk as this CEP unleashes clear states as a node progresses through a restart. Accounting for a variety of edge cases, an operator can feel assured that, for example, a non-functioning sidecar won’t derail operations.
The current process for restarting a node is a multi-step, manual process, which does not scale for large cluster sizes (and is also tedious for small clusters). Restarting clusters previously lacked a streamlined approach as each node needed to be restarted one at a time, making the process time-intensive and error-prone.
Though Sidecar is still considered WIP, it’s got big plans to improve operating large clusters!
The NetApp Instaclustr Platform, in conjunction with our expert TechOps team already orchestrates these laborious tasks for our Cassandra customers with a high level of care to ensure their cluster stays online. Restarting or upgrading your Cassandra nodes is a huge pain-point for operators, but it doesn’t have to be when using our managed platform (with round-the-clock support!)
CEP-54: Zstd with dictionary SSTable compressionCEP-54, with NetApp Instaclustr’s collaboration, aims to add support Zstd with dictionary compression for SSTables. Zstd, or Zstandard, is a fast, lossless data compression algorithm that boasts impressive ratio and speed and has been supported in Cassandra since 4.0. Certain workloads can benefit from significantly faster read/write performance, reduced storage footprint, and increased storage device lifetime when using dictionary compression.
At a high level, operators choose a table they want to compress with a dictionary. A dictionary must be trained first on a small amount of already present data (recommended no more than 10MiB). The result of a training is a dictionary, which is stored cluster-wide for all other nodes to use, and this dictionary is used for all subsequent writes of SSTables to a disk.
Workloads with structured data of similar rows benefit most from Zstd with dictionary compression. Some examples of ideal workloads include event logs, telemetry data, metadata tables with templated messages. Think: repeated row data. If the table data is too unstructured or random, this feature likely won’t be optimal for dictionary compression, however plain Zstd will still be an excellent option.
New SSTables with dictionaries are readable across nodes and can stream, repair, and backup. Existing tables are unaffected if dictionary compression is not enabled. Too many unique dictionaries hurt decompression; use minimal dictionaries (recommended dictionary size is about 100KiB and one dictionary per table) and only adopt new ones when they’re noticeably better.
CEP-55: Generated role namesCEP-55 adds support to create users/roles without supplying a name, simplifying
user management, especially when generating users and roles in bulk. With an example syntax, CREATE GENERATED ROLE WITH GENERATED PASSWORD, new keys are placed in a newly introduced configuration section in cassandra.yaml under “role_name_policy.”
Stefan Miklosovic, our Cassandra engineer at NetApp Instaclustr, created this CEP as a logical follow up to CEP-24 (password validation/generation), which he authored as well. These quality-of-life improvements let operators spend less time doing trivial tasks with high-risk potential and more time on truly complex matters.
Manual name selection seems trivial until a hundred role names need to be generated; now there is a security risk if the new usernames—or worse passwords—are easily guessable. With CEP-55, the generated role name will be UUID-like, with optional prefix/suffix and size hints, however a pluggable policy is available to generate and validate names as well. This is an opt-in feature with no effect to the existing method of generating role names.
The future of Apache Cassandra is brightThese Cassandra Enhancement Proposals demonstrate a strong commitment to making Apache Cassandra more powerful, secure, and easier to operate. By staying on top of these updates, we ensure our managed platform seamlessly supports future releases that accelerate your business needs.
At NetApp Instaclustr, our expert TechOps team already orchestrates laborious tasks like restarts and upgrades for our Apache Cassandra customers, ensuring their clusters stay online. Our platform handles the complexity so you can get up and running fast.
Learn more about our fully managed and hosted Apache Cassandra offering and try it for free today!
The post Stay ahead with Apache Cassandra®: 2025 CEP highlights appeared first on Instaclustr.
Vector search benchmarking: Embeddings, insertion, and searching documents with ClickHouse® and Apache Cassandra®
Welcome back to our series on vector search benchmarking. In part 1, we dove into setting up a benchmarking project and explored how to implement vector search in PostgreSQL from the example code in GitHub. We saw how a hands-on project with students from Northeastern University provided a real-world testing ground for Retrieval-Augmented Generation (RAG) pipelines.
Now, we’re continuing our journey by exploring two more powerful open source technologies: ClickHouse and Apache Cassandra. Both handle vector data differently and understanding their methods is key to effective vector search benchmarking. Using the same student project as our guide, this post will examine the code for embedding, inserting, and retrieving data to see how these technologies stack up.
Let’s get started.
Vector search benchmarking with ClickHouseClickHouse is a column-oriented database management system known for its incredible speed in analytical queries. It’s no surprise that it has also embraced vector search. Let’s see how the student project team implemented and benchmarked the core components.
Step 1: Embedding and inserting data
scripts/vectorize_and_upload.py
This is the file that handles Step 1 of the pipeline for
ClickHouse. Embeddings in this file
(scripts/vectorize_and_upload.py) are used as vector
representations of Guardian news articles for the purpose of
storing them in a database and performing semantic search. Here’s
how embeddings are handled step-by-step (the steps look similar to
PostgreSQL).
First up, is the generation of embeddings. The same
SentenceTransformer model used in part 1
(all-MiniLM-L6-v2) is loaded in the class constructor.
In the method generate_embeddings(self, articles), for
each article:
- The article’s title and body are concatenated into a text string.
- The model generates an embedding vector
(
self.model.encode(text_for_embedding)), which is a numerical representation of the article’s semantic content. - The embedding is added to the article’s dictionary under the
key
embedding.
Then the embeddings are stored in ClickHouse as follows.
- The database table
guardian_articlesis created with an embeddingArray(Float64) NOT NULLcolumn specifically to store these vectors. - In
upload_to_clickhouse_debug(self, articles_with_embeddings), the script inserts articles into ClickHouse, including the embedding vector as part of each row.
services/clickhouse/clickhouse_dao.py
The steps to search are the same as for PostgreSQL in part 1.
Here’s part of the related_articles method for
ClickHouse:
def related_articles(self, query: str, limit: int =
5):
"""Search for similar articles using vector similarity""" ... query_embedding = self.model.encode(query).tolist() search_query = f""" SELECT url, title, body, publication_date, cosineDistance(embedding, {query_embedding}) as distance FROM guardian_articles ORDER BY distance ASC LIMIT {limit} """ ...
When searching for related articles, it encodes the query into an embedding, then performs a vector similarity search in ClickHouse using cosineDistance between stored embeddings and the query embedding, and results are ordered by similarity, returning the most relevant articles.
Vector search benchmarking with Apache CassandraNext, let’s turn our attention to Apache Cassandra. As a distributed NoSQL database, Cassandra is designed for high availability and scalability, making it an intriguing option for large-scale RAG applications.
Step 1: Embedding and inserting datascripts/pull_docs_cassandra.py
As in the above examples, embeddings in this file are used to
convert article text (body) into numerical vector
representations for storage and later retrieval in Cassandra.
For each article, the code extracts the body and
computes the embeddings:
embedding = model.encode(body) embedding_list = [float(x) for x in embedding]
model.encode(body)converts the text to aNumPyarray of 384 floats.- The array is converted to a standard Python list of floats for Cassandra storage.
Next, the embedding is stored in the vector column of the
articles table using a CQL INSERT:
insert_cql = SimpleStatement(""" INSERT INTO articles (url, title, body, publication_date, vector) VALUES (%s, %s, %s, %s, %s) IF NOT EXISTS; """) result = session.execute(insert_cql, (url, title, body, publication_date, embedding_list))
The schema for the table specifies: vector
vector<float, 384>, meaning each article has a
corresponding 384-dimensional embedding. The code also creates a
custom index for the vector column:
session.execute(""" CREATE CUSTOM INDEX IF NOT EXISTS ann_index ON articles(vector) USING 'StorageAttachedIndex'; """)
This enables efficient vector (ANN: Approximate Nearest Neighbor) search capabilities, allowing similarity queries on stored embeddings.
A key part of the setup is the schema and indexing. The
Cassandra schema in
services/cassandra/init/01-schema.cql defines the
vector column.
Being a NoSQL database, Cassandra schemas are a bit different to normal SQL databases, so it’s worth taking a closer look. This Cassandra schema is designed to support Retrieval-Augmented Generation (RAG) architectures, which combine information retrieval with generative models to answer queries using both stored data and generative AI. Here’s how the schema supports RAG:
- Keyspace and table structure
- Keyspace (
vectorembeds): Analogous to a database, this isolates all RAG-related tables and data. - Table (
articles): Stores retrievable knowledge sources (e.g., articles) for use in generation.
- Keyspace (
- Table columns
url TEXT PRIMARY KEY: Uniquely identifies each article/document, useful for referencing and deduplication.title TEXTandbody TEXT: Store the actual content and metadata, which may be retrieved and passed to the generative model during RAG.publication_date TIMESTAMP: Enables filtering or ranking based on recency.vector VECTOR<FLOAT, 384>: Stores the embedding representation of the article. The new Cassandra vector data type is documented here.
- Indexing
- Sets up an Approximate Nearest Neighbor (ANN) index using Cassandra’s Storage Attached Index.
More information about Cassandra vector support is in the documentation.
Step 2: Vector search and retrievalThe retrieval logic in
services/cassandra/cassandra_dao.py showcases the
elegance of Cassandra’s vector search capabilities.
The code to create the query embeddings and perform the query is similar to the previous examples, but the CQL query to retrieve similar documents looks like this:
query_cql = """ SELECT url, title, body, publication_date FROM articles ORDER BY vector ANN OF ? LIMIT ? """ prepared = self.client.prepare(query_cql) rows = self.client.execute(prepared, (emb, limit))What have we learned?
By exploring the code from this RAG benchmarking project we’ve seen distinct approaches to vector search. Here’s a summary of key takeaways:
- Critical steps in the process:
- Step 1: Embedding articles and inserting them into the vector databases.
- Step 2: Embedding queries and retrieving relevant articles from the database.
- Key design pattern:
- The DAO (Data Access Object) design pattern provides a clean, scalable way to support multiple databases.
- This approach could extend to other databases, such as OpenSearch, in the future.
- Additional insights:
- It’s possible to perform vector searches over the latest documents, pre-empting queries, and potentially speeding up the pipeline.
So far, we have only scratched the surface. The students built a complete benchmarking application with a GUI (using Steamlit), used multiple other interesting components (e.g. LangChain, LangGraph, FastAPI and uvicorn), Grafana and LangSmith for metrics, and Claude to use the retrieved articles to answer questions, and Docker support for the components. They also revealed some preliminary performance results! Here’s what the final system looked like (this and the previous blog focused on the bottom boxes only).
In a future article, we will examine the rest of the application code, look at the preliminary performance results the students uncovered, and discuss what they tell us about the trade-offs between these different databases.
Ready to learn more right now? We have a wealth of resources on vector search. You can explore our blogs on ClickHouse vector search and Apache Cassandra Vector Search (here, here, and here) to deepen your understanding.
The post Vector search benchmarking: Embeddings, insertion, and searching documents with ClickHouse® and Apache Cassandra® appeared first on Instaclustr.
Optimizing Cassandra Repair for Higher Node Density
This is the fourth post in my series on improving the cost efficiency of Apache Cassandra through increased node density. In the last post, we explored compaction strategies, specifically the new UnifiedCompactionStrategy (UCS) which appeared in Cassandra 5.
- Streaming Throughput
- Compaction Throughput and Strategies
- Repair (you are here)
- Query Throughput
- Garbage Collection and Memory Management
- Efficient Disk Access
- Compression Performance and Ratio
- Linearly Scaling Subsystems with CPU Core Count and Memory
Now, we’ll tackle another aspect of Cassandra operations that directly impacts how much data you can efficiently store per node: repair. Having worked with repairs across hundreds of clusters since 2012, I’ve developed strong opinions on what works and what doesn’t when you’re pushing the limits of node density.
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 zonesNetApp 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 clustersAs 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 strategyTo 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 successSome 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 CassandraTo 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 ProxyWe 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
Across different cluster sizes, we observed no significant difference in operation throughput between PrivateLink and non-PrivateLink configurations.
Latency resultsLatency 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 77.1 141.2Results 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 todayThe 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.
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...