Apache Cassandra is an open-source, distributed NoSQL database widely adopted by companies like Netflix, eBay, Apple, and more. However, due to unpredictable and sluggish Cassandra performance, finicky manual tuning, very slow release cycles, and lack of critical features, many companies have made the switch from Cassandra to Scylla.
Scylla acts as an Apache Cassandra alternative and replacement with fully supported Cassandra migration capabilities. Scylla provides the same CQL interface and queries, the same drivers, even the same on-disk SSTable format, but with a modern architecture designed to eliminate Cassandra performance issues, limitations, and operational barriers.
For these reasons and more, Scylla users have realized faster, more consistent performance for their mission-critical applications, saved millions in infrastructure costs and freed up countless hours previously spent tuning their systems in attempts to get desired levels of performance. In this Scylla vs Cassandra breakdown, we explore Scylla’s advantages over Cassandra, followed by a comprehensive Cassandra features comparison.
The total cost of ownership (TCO) with Scylla is much less than that of Cassandra. Scylla was designed from the ground-up to deliver the best possible price-performance. Its low-level design squeezes every cycle from your CPU. Scylla is designed to run as close as possible to 100% CPU utilization, with every operation given a priority class. There is no need to wastefully over-provision.
Scylla delivers consistent, reliable real-time performance. That’s because its built-in schedulers prioritize reads and writes over maintenance tasks, such as repairs and compactions — eliminating latency spikes. With Scylla, you never experience garbage collection stalls, which impact Cassandra performance. To avoid such interruptions, Scylla adopts a comprehensive ‘shared-nothing’ design. No locks are taken, so latency is never affected.
GE Predix was able to greatly reduce the administrative burden to meet their SLAs after switching to Scylla.
Speed doesn’t have to come at the price of complexity. Scylla simplifies everything for its users. It automatically configures the RAID device for you with the right striping and automatically assigns the NICs network queues to shards. Scylla installs daemons in an isolated Linux control group to cap their memory/CPU usage. Scylla’s setup tool runs a disk benchmark to maximize throughput while keeping latency low. And with Scylla, DevOps can leave tuning to the database itself.
Expedia was able to reduce P99 response times from >80ms to <5ms after migrating to Scylla NoSQL.
Scylla’s innovative shard-per-core design divides a server’s resources into shared-nothing units of CPU-core, RAM, persistent storage and network I/O. Scylla runs at near maximum utilization on all available cores of multi-CPU hardware architectures. Scylla’s end-to-end sharded architecture, along with shard-aware drivers means that each client writing or requesting data can send queries directly to the CPU core responsible for that shard of data. This minimizes hot shards and removes extra hops.
At the heart of Scylla lies its core engine, the Seastar framework, a standalone library developed by ScyllaDB. Seastar has a specialized scheduler with a fully async programming paradigm of futures and promises and advanced concepts like coroutines, that can run a million lambda functions per core per second. Seastar is responsible for scheduling, networking, Direct Memory Access, shard-per-core memory allocators, and so forth. Seastar is written in C++20 and uses every innovative trick and paradigm. Find out more about Scylla’s “everything-async” architecture here.
Scylla uses C++20 and the best compiler techniques, as well as deep knowledge of the Linux kernel and per-core hyperthreading to maximize CPU utilization. Scylla also automatically configures your network card interrupts to balance interrupt request (IRQ) processing across CPU cores. Scylla explicitly chooses to read-ahead data from the drive when it expects a follow-on disk access instead of blindly relying on the disk, as is the case with Cassandra. Scylla controls all aspects of CPU execution and runs procedures to use the CPU idle time so memory layout will be optimized.
Like Apache Cassandra, Scylla’s client driver is topology-aware and will prefer a node that owns the key range under query. Scylla takes this design one step further and enables the client to reach the specific CPU core within the replica that owns that shard of data. This design improves load balancing among the servers and provides superior performance by avoiding needless concurrency overhead. Shard-aware Scylla drivers are available for Java, Go, Python, C++ and Rust. However, Scylla also fully supports standard Cassandra/DSE drivers for frictionless replacement compatibility.
Scylla’s lightweight transactions (LWT) use a Paxos algorithm mechanism similar to Cassandra, but they involve one fewer round trips, making them more efficient, faster, and with lower latency. While Cassandra issues a separate read query to fetch the old record, Scylla piggybacks the read result on the response to the prepare round. Scylla’s LWT has a special commitlog mode that automatically balances between the transaction durability flush requirement and fast, non-transactional operations.
Per-query cache bypass clause on SELECT statements allows for range scan queries that typically process large amounts of data to skip reading the cache and avoids populating the cache, flooding it with data. Bypass cache keeps your working set in-cache, enabling you to squeeze more performance from your cluster and minimizes RAM overhead, so real-time queries receive the best latency.
Scylla is designed to be Non-Uniform Memory Access (NUMA)-friendly, with highly optimized memory management down to the application binaries. Each shard of data is assigned its own chunk of RAM and CPU core, bound within the same socket. This is referred to as “NUMA-local” processing. A non-NUMA-friendly deployment causing memory access to be twice as expensive.
Scylla implements a different repair checksum algorithm that resembles rsync and runs the checksum at row granularity instead of partition granularity. The new algorithm conciliates repair faster and more efficiently, sends less data over the wire reducing network traffic and is less sensitive to large partitions.
DSE instead uses a Merkle tree mechanism for anti-entropy repairs. It has no row-level repair mechanism, and thus repairs take longer and are far less efficient, generating many times more network traffic and degrading overall performance.
Scylla’s I/O scheduler prioritizes read/write operations class over compactions. When reads and writes spike, Scylla automatically queues compaction activity so that they do not impact your throughput or latencies. Scylla runs compaction at full speed only when there is CPU/disk idle time, with all cores running compactions in parallel.
With Cassandra and its heavier utilization of system resources, you have to worry about and track compactions. When making a Cassandra comparison keep in mind that Scylla requires no brittle tuning.
Scylla allows for operational and analytics workloads to run against a shared cluster — a unique feature we call Workload Prioritization. Scylla’s built-in scheduler prioritizes transactions and tasks based on ‘shares’ of system resources assigned per-user, balancing requests to maintain desired service level agreements (SLAs) for each service. This enables you to run a single cluster that is scaled to support both types of operations, simplifying your architecture and saving on hardware provisioning.
With Cassandra, you have no control over system utilization so that, for example, full scan queries could cause timeouts, blocking ongoing write transactions.
Scylla is written in natively compiled C++, and thus does not use a Java Virtual Machine (JVM). There is no need to optimize JVM flags, and you can forget about Garbage Collection (GC) stalls, tuning, and surprises. With Scylla, there is no need to compute heap sizes, no need to divide the RAM between the JVM, the off-heap, and the page cache. Cassandra suffers from the worst of all possible worlds, having to manage memory (pools, off-heap), ongoing tuning, and suffering slowdowns due to the JVM.
Forget about tuning your database! Upon installation, Scylla runs a benchmark to measure your disk and will make all of the Linux configurations on your behalf — from RAID setup to clock drift and fstrim disk scheduling. Using control theory, Scylla continuously makes the database less fragile by dynamically tuning the way resources are used instead of requiring an operator to adjust an overwhelming number of configurations on the fly.
With Cassandra, you have to tune many aspects manually, including the JVM, bloom filters, row and key caches, and even memtable thresholds. Not to mention all the operating system and hardware settings.
Scylla has a unified row-based cache system that automatically tunes itself, allowing it to adapt to different data access patterns and workloads. With Scylla, with its inherent low-latency design, there’s no need for external caches, further simplifying the infrastructure.
Cassandra uses several separate competing caches (key cache, row cache, on/off-heap, and Linux OS page cache) that require an operator to analyze and tune — a manual process that will never be able to keep up with users’ dynamic workloads.
Scylla adopts open standards and allows you to use the best open-source tools. Scylla’s metrics are based on Prometheus for collection, Grafana dashboards for presentation, and Grafana Loki for log aggregation. Both Prometheus and Grafana are Cloud Native Computing Foundation (CNCF) graduated projects. Scylla contributed to Wireshark to add support for CQL and also for its internal RPC for better traceability. Scylla strives to maintain Apache Cassandra compatibility.
Comparing Scylla vs Cassandra, it becomes clear that Scylla provides flexibility, supporting both local and global secondary indexes. Scylla allows for tables to have global secondary indexes across all nodes in a cluster, not just locally on a single node. This is an efficient way to lookup rows because you can find the node hosting the row by hashing the partition key. However, this also means that finding a row using a non-partition key requires a full table scan which is inefficient. That’s why we also support local secondary indexes.
Cassandra and other Cassandra alternatives support only local indexes, which aren’t scalable and aren’t designed for performing full table scans.
ScyllaDB employs the best open-source experts and has a legacy of consistent open-source contributions. We are fully committed to open source. Seastar, Scylla’s core engine, is used at the core of other open-source projects such as Redpanda data streaming engine, the Ceph storage engine, and others. The shard-aware GoCQLX, Python and Rust drivers were developed by the ScyllaDB team, with contributions from the open-source community. ScyllaDB has also offered contributions to the broader open source community, such as enhancements to Linux XFS in order to make it more asynchronous, contributed kernel code for system call efficiency, and made numerous other contributions.
Starting with Scylla 4.0, node operations such as streaming and decommission are based on repair algorithms under the hood. This enables you to pause or restart repair operations while going back to the same checkpoint position, saving a lot of time on administrative operations and keeping repair impact as low as possible on your cluster.
Cassandra has no such feature, but offers NodeSync for continuous background repairs. However, DataStax admits, “It is not a race-free lock; there is a possibility of duplicated work.”
Scylla enables you to easily and consistently track and stream table changes. Change data is stored in Scylla as a standard CDC-readable table that developers can query using standard CQL. The data is consistent across the replica set and can provide the previous version of the data changed. Cassandra CDC has been introduced with version 4.0, however, Scylla’s CDC surpasses Cassandras’ CDC in terms of ease of use and functionality. It supports capturing the pre- and post-image as well as the delta (specific changes) for the record. Query results are automatically de-duplicated.
Cassandra 4.0, on the other hand, stores CDC in a special commitlog-like structure on each node. It cannot be read using CQL queries, and requires special applications to be written to aggregate and de-duplicate the data.
Scylla enables you to scale up and out efficiently. Its ability to fully utilize system resources allows Scylla to run on smaller clusters of larger nodes. Scylla is adding Raft consensus protocol to permit scaling out of multiple nodes at a time.
When the time comes to scale your Cassandra deployment, you will find that Cassandra can add only one node at a time. Plus, since Cassandra nodes will generally be smaller instances, each addition only brings on a lower amount of storage and compute. As a result, it will take at least 10x longer to expand your cluster. Since expanding your cluster hinders your ability to quickly react to changes, you’re forced to overprovision up-front.
Scylla supports Amazon DynamoDB-compatible operations, including a version of DynamoDB Streams, which we call Project Alternator. This provides users more flexibility in data models and query APIs. Scylla allows you to avoid vendor lock-in. You are free to choose among multiple database APIs and at any point change your physical deployment, from on-premises to the cloud vendor of your choice.
Cassandra provides only a CQL-compatible API. It offers no DynamoDB compatibility.
Heat-weighted load balancing efficiently performs rolling node upgrades and reboots by allowing cold nodes to slowly ramp up into requests as its cache is being populated. This aids after restarts or any other time when a cluster loses its cache.
Cassandra does not support heat-weighted load balancing.
Scylla can linearly scale-up performance, even on the largest machines available today – such as the AWS i3en.24xlarge, which provides 60TB of local NVMe SSD storage. Custom-built on-premises equipment can scale even larger. On Scylla, compaction and streaming on such a large instance takes exactly the same amount of time as on a small i3en.xlarge.
Cassandra, written in Java and limited by JVM performance, has issues with nodes larger than 2TB. Exceeding this size per node may result in delays in bootstrapping, repairs, compactions, and recovery.
Scylla’s incremental compaction strategy (ICS) enhances Cassandra’s existing Size-tiered Compaction Strategy (STCS) strategy by dividing SStables into increments. ICS greatly reduces the temporary space amplification which is typical of STCS, resulting in more disk space being available for storing user data, eliminating the typical requirement of 50% free space in your drive.
There is no comparable Cassandra feature.
See Scylla vs Apache Cassandra, Amazon DynamoDB and Google BigTable