The 10 best NoSQL cloud databases

You may have seen our two earlier posts in this three-post series: NoSQL cloud databases: Benefits and features explained, and NoSQL cloud database use case guide. This is the last post, which helps you determine how to pick one.

A lot has happened since we first published this guide back in 2024. Two of the ten products changed owners: IBM completed its acquisition of DataStax in May 2025, and Haveli Investments took Couchbase private in September 2025. Two changed licensing: Redis added AGPLv3 alongside its source-available licenses with Redis 8, and ScyllaDB moved to a source-available license entirely. And vector search, which barely featured in operational databases two years ago, now ships natively in seven of the ten. Those all changed how to evaluate a shortlist, so this version adds ownership, licensing, and vector capability alongside the usual questions about consistency, scale, and cost.

Every product here is a legitimate choice for some workload. None is the right answer for all of them. The comparison table below is the fastest way to narrow the field; sections that follow explain what each one is good at and where it isn’t. Every factual claim is sourced to vendor documentation or announcements, and where a vendor's own docs name a limit, we have used their words rather than ours.

Here are the ten, in alphabetical order.

The 10 NoSQL cloud databases at a glance

Aerospike Cloud

Aerospike Cloud is a fully managed DBaaS built for predictable performance as data volume grows, scaling alongside your applications from gigabytes to petabytes. It is designed for high-throughput, low-latency applications and deploys with cross-zone replication, so a cluster survives the loss of an availability zone. The degree of protection depends on the topology, which ranges from replication factor 2 (RF2) across two zones to replication factor 3 (RF3) across three. Every cluster includes an SLA, which varies by configuration.¹

Why it stands out:

• Proven at scale: Dream11 runs Aerospike on AWS, adding 35 TB per day at millions of transactions per second with sustained sub-15ms latency.^{1, 3}

• Predictable performance: Hybrid Memory Architecture is the default storage configuration, holding the primary index in memory and record data on SSD. In-memory and all-flash configurations are also available, and each namespace can be configured separately.²

• No transaction charges: Unlike the hyperscaler databases on this list, Aerospike Cloud prices compute, cloud management, and network transfer, not requests.¹

Top features:

• Distributed ACID transactions with strict serializability, introduced in Aerospike Database 8⁴

• Automatic failover, rebalancing, and reconfiguration, with no dropped traffic during those operations¹

• Cluster provisioning and management through both a console UI and an automation API¹

• Multi-model support for key-value and document/JSON APIs¹

• Selectable consistency mode per cluster, high availability (AP) or strong consistency (SC)¹

Benefits:

• Failovers, rebalancing, and upgrades are handled by the service, with 24/7 support from Aerospike engineers and integrations into existing observability tooling, including Datadog and Grafana.¹

• SOC 2 certified, with Type II attestation available on request. Traffic stays inside VPC isolation and private peering, and encryption at rest and in transit is on by default.¹

• Deployment is not locked to one cloud. Aerospike Cloud runs on AWS and Google Cloud, and the underlying database can also be self-managed in your own data center, so the platform choice does not have to be permanent.¹

Cons:

• Parts of the developer toolchain are still pre-release. The Developer SDK is in preview, with the Python SDK at an alpha release⁵

• Pricing is capacity-based, so you pay for provisioned compute and storage rather than actual request volume, which favors steady workloads over spiky ones¹

• Google Cloud support is not at parity with AWS, because self-service deployment on GCP has not yet shipped¹

Best for: Businesses that aim for high-performance applications, such as AdTech, that require real-time data processing at a global scale.

Organizations that prioritize predictable latency under load will find the storage model and per-cluster consistency controls do most of the work here. Aerospike Cloud fits critical applications such as fraud detection and customer engagement platforms, and serves as a real-time user profile database.

Amazon DynamoDB

If your organization uses Amazon Web Services (AWS) as its cloud provider, it makes sense to consider Amazon DynamoDB, Amazon's NoSQL cloud-managed services product.

Why it stands out:

• Proven at extreme scale: Over the four days of Prime Day 2025, Amazon properties made tens of trillions of calls to the DynamoDB API, with DynamoDB peaking at 151 million requests per second while maintaining single-digit millisecond responses.⁶

• Zero-RPO multi-region strong consistency: Global tables can be configured so every replica always returns the latest write, a guarantee few managed NoSQL services offer. It reached general availability in June 2025.^{7, 8}

• Vector search alongside operational data: Native vector indexing reached general availability in August 2026, so embeddings are stored in the same table as the records they describe, with no separate vector store to sync.⁹

Top features:

• Global tables with two consistency modes: multi-region eventual consistency by default, or multi-region strong consistency across three regions⁸

• On-demand and provisioned capacity modes, with no servers to provision or patch in either¹¹

• Enterprise-grade security with fine-grained access control¹¹

• DynamoDB Streams for change data capture, feeding downstream processing and replication¹¹

• ACID transactions spanning multiple items and tables within a region¹¹

Benefits:

• The same table design scales from a low-traffic application to Prime Day volumes without re-designing, which removes a migration most teams otherwise face.⁶

• Fully serverless operation with automated backups and point-in-time recovery, so there is nothing to provision, patch, or size.¹¹

• Native integration across the AWS stack, including Bedrock for generating the embeddings that vector indexes store, and IAM for access control.^{9, 11}

Cons:

• A hard 400 KB per-item size limit, which forces access patterns to be designed up front and rules out storing large documents or blobs in the table¹⁰

• Request-based pricing makes cost forecasting difficult, because cost tracks read and write volume rather than provisioned footprint, and two workloads with the same monthly totals cost differently depending on the spikiness of the traffic¹¹

• Multi-region strong consistency comes with real constraints: it requires exactly three regions, does not support the transaction APIs, and a global table's consistency mode cannot be changed after creation⁸

Best for: Companies of all sizes that require a database with good performance and scalability, particularly those already integrated within the AWS ecosystem.

The case for DynamoDB is rarely the database in isolation. It is the fit with everything else in an AWS account, from IAM to Bedrock, and the operational model that comes with a service Amazon runs at its own scale. For teams already committed to AWS, that integration is usually the deciding factor.

Couchbase Capella

Couchbase Capella offers more flexibility for organizations that run on more than one cloud provider, or that want to keep their cloud provider options open for the future. Couchbase went private in September 2025 when Haveli Investments completed a roughly $1.5 billion all-cash acquisition, so roadmap and pricing decisions are now made by a private equity owner rather than public markets.¹²

Why it stands out:

• Deploy on any major cloud: Capella runs on AWS, Google Cloud, and Microsoft Azure, and the same engine can be self-managed on-premises, so a move between providers is not a re-platform.¹³

• Search across cloud and edge: Couchbase ships vector search across Capella, self-managed Server, and Couchbase Lite. Server 8.0 added Hyperscale and Composite vector indexes in October 2025, alongside the existing Search Vector indexes for hybrid queries that combine vector similarity with text or geospatial search.^{14, 15}

• Couchbase Lite: An embeddable database for mobile, Internet of Things (IoT), and web apps, with cloud-to-device and peer-to-peer sync, deterministic version-vector conflict resolution, and on-device vector search that runs semantic queries with no network connection.^{25, 26}

Top features:

• Queries JSON data with SQL++ (formerly N1QL), a complete SQL dialect familiar to anyone with relational experience¹⁶

• Full-text, geospatial, and vector search in one engine, all reachable from one SQL++ query^{14, 15}

• Includes fully managed integrated cache layer¹⁷

• Capella Columnar for real-time JSON analytics alongside operational workloads, generally available since September 2024¹⁸

• Memory-first architecture, with the Magma storage engine as the default for datasets that exceed available memory^{17, 19}

Benefits:

• Couchbase Capella supports familiar SQL-like queries, making it easier for developers who need to transition from relational databases.¹⁶

• A cloud-agnostic setup avoids vendor lock-in, and the same engine can run self-managed if cloud is not an option.¹³

• The memory-first design serves the working set from cache, while Magma handles datasets larger than memory at a memory-to-data ratio as low as 1%.^{17, 19, 20}

Cons:

• Magma's performance is tied to the underlying disk subsystem. Couchbase recommends SSDs for the data path, so hardware quality is a direct input to performance rather than something you can trade away^{19, 21}

• ACID transactions are scoped to a single cluster. Cross Data Center Replication is asynchronous and eventually consistent, and Couchbase advises against combining transactions with active-active bidirectional replication, so multi-region deployments cannot get transactional guarantees across regions^{22, 23}

• The platform has broadened and been renamed repeatedly. Columnar reached general availability in September 2024, and the AI offering went from Capella AI Services in private preview in December 2024, to general availability in December 2025, to the current AI Data Plane. Each addition is more surface area to evaluate, price, and keep current with, and renamed products make older documentation and community answers harder to trust^{18, 24, 27}

Best for: Organizations that like Couchbase’s SQL++ query language and that wish to take advantage of its mobile sync capabilities.

Couchbase is the only entry with a serious offline story, so the question is rarely Capella against another cloud database in isolation. It is whether an application needs the same query language and sync model running in a data center and on a device that loses connectivity, a narrow requirement that almost nothing else satisfies.

DataStax Astra DB (now IBM)

DataStax Astra DB has changed hands since this guide was first published. IBM completed its acquisition of DataStax on May 28, 2025, and Astra DB is now sold as a component of the IBM watsonx.data multicloud offering, where it extends a lakehouse architecture with a fully managed, cloud-native NoSQL service.^{28, 29, 30} The Astra Classic product was rebranded as Astra Managed Clusters.³¹ The underlying engine is still Apache Cassandra, so the technical profile is largely unchanged from 2024. What has changed is where you buy it and how it is packaged. Like Couchbase Capella, Astra DB is not tied to one cloud provider, which suits organizations running hybrid services or wanting to switch providers more easily.

Why it stands out:

• Cassandra lineage: Astra DB runs Apache Cassandra, so existing Cassandra Query Language (CQL), drivers, and Cassandra operational knowledge carry over.³³

• Vector search on an operational engine: Astra DB is intended for generative AI workloads, with vector search running against the same Cassandra tables that serve operational traffic rather than a separate store.³²

• Embedded in an enterprise AI stack: Astra DB is configurable from the watsonx.data Infrastructure Manager, so it is provisioned and governed alongside a lakehouse rather than as a standalone database.³⁰

Top features:

• Supports wide-column, key-value, time series, and vector data models^{32, 33}

• Cloud-agnostic deployment across the major public clouds, with multi-region databases supported³³

• Two access paths: a Data API for document-style development and CQL through the shell and Cassandra drivers³³

• Triple replication managed for you, with data replicated across three availability zones in every deployed region at a replication factor of 3³³

• Configurable consistency levels through CQL, including EACH_QUORUM, SERIAL, and ALL for writes that must be acknowledged across all regions^{33, 34}

Benefits:

• Cassandra compatibility means CQL, existing drivers, and Cassandra operational experience transfer without retraining.³³

• Cloud-agnostic support works well for hybrid organizations that want to keep their cloud options open.³³

• Vector and operational data are stored in the same tables, and IBM's watsonx integration puts governance and access control in the same place as the rest of an enterprise AI stack.^{30, 32}

Cons:

• Astra DB follows an eventual consistency model, and replicas cannot be independently managed, restricted, or modified. The Data API and clients always use LOCAL_QUORUM, and write consistency levels ONE, ANY, and LOCAL_ONE are unavailable^{33, 34}

• Cross-region replication is asynchronous. DataStax's documentation says replication across regions generally takes minutes but can take hours or days in the event of failures or outages³⁵

• The legacy Document, REST, GraphQL, and gRPC APIs for Astra Managed Clusters reached end of life in May 2026, so applications built on them need to be migrated to the Data API or CQL³¹

Best for: Distributed data environments requiring only eventual consistency.

Astra DB is the clearest case on this list where the acquisition matters more than the engine. Cassandra's strengths and limits are well understood and largely unchanged. The question a buyer faces in 2026 is whether being inside IBM's watsonx portfolio is an advantage, because it brings lakehouse governance and enterprise AI tooling, or a constraint, because the database's roadmap is now set by a much larger platform strategy.

Google Cloud Bigtable

Google Cloud Bigtable is the managed version of the wide-column store Google has run internally for more than 20 years, first described in a 2006 research paper and still underpinning Search, Ads, Maps, and YouTube.^{36, 44} It is a sparse, distributed key-value store indexed by row key, which makes it a different product from BigQuery, Google's analytics warehouse, despite the similar name.⁴¹ Google positions it as the NoSQL pioneer, and scalability is where that lineage shows.³⁷

Why it stands out:

• Massive scalability: Bigtable is the system behind Google Search, Ads, Maps, and YouTube. As of April 2024, Google reported it serving 7 billion queries per second at peak across its fleet, with more than 10 exabytes under management.³⁶ For one customer deployment, capacity is provisioned per node, with each node serving up to 17,000 requests per second.³⁷

• Wide-column with cell versioning: Each intersection of a row and column can hold multiple cells, each a uniquely timestamped version of the value, so the data model records how a value changed over time without a separate history table.⁴¹

• SQL on a NoSQL engine: Bigtable's SQL interface reached general availability in April 2025, and continuous materialized views let you define an aggregation in SQL that Bigtable keeps current incrementally, without affecting read and write performance.^{37, 38}

Top features:

• Easy administration for tasks such as replication and cluster resizing³⁷

• Automatic load balancing helps keep response times steady³⁷

• Multi-region replication with configurable single-cluster or multi-cluster routing policies⁴⁰

• GoogleSQL interface, generally available since April 2025, runnable from Bigtable Studio or programmatically through the Java, Python, and Go client libraries^{38, 43}

• Continuous materialized views that update incrementally as data arrives, so real-time aggregations do not require a separate pipeline^{37, 38}

Benefits:

• The availability SLA scales with topology, reaching 99.999% for replicated instances using multi-cluster routing across three or more regions.³⁹

• Customer-managed encryption keys with external key manager support, IAM integration, VPC Service Controls, and audit logging meet the controls most regulated industries require, with access authorized down to the table, column, or row level.^{37, 42}

• Integration with Dataflow, Dataproc, and BigQuery keeps Bigtable inside an existing Google Cloud data stack.³⁷

Cons:

• Cost climbs quickly at sustained high volume, because you pay for provisioned nodes rather than per request³⁷

• SQL support is read-only. There is no DML beyond SELECT, no DDL, and no subqueries, JOINs, UNIONs, or CTEs, so it is a query convenience rather than a full query engine. SQL requests also do not fail over to another cluster, even under a multi-cluster routing policy^{38, 40, 43}

• Availability and transactions are mutually exclusive. Bigtable prevents you from enabling single-row transactions on an app profile that uses multi-cluster routing, because there is no safe way to run both, so multi-cluster availability means giving up read-modify-write operations⁴⁰

Best for: Developers building applications that need high throughput and scalability for key-value data, such as customer 360, financial services, and IoT.

If you're concerned about scalability, a NoSQL database that supports Google's operations can probably handle yours.

Microsoft Azure Cosmos DB

If your organization is already using Microsoft Azure, it makes sense to look at Microsoft's NoSQL offering, Azure Cosmos DB. One engine stores JSON items natively and exposes them through six APIs, one proprietary and five compatible with the wire protocols of MongoDB, Cassandra, Gremlin, Azure Table Storage, etc.⁵¹ Cosmos DB is not several databases but one, wearing whichever interface your application already speaks.

Why it stands out:

• Five consistency levels, not two: Microsoft's documentation notes that most commercially available distributed NoSQL databases offer only strong and eventual consistency. Cosmos DB offers strong, bounded staleness, session, consistent prefix, and eventual consistency, configurable per account or overridden per request.⁴⁵

• SLAs on four dimensions: Azure backs throughput, latency, availability, and consistency contractually, including latency at the 99th percentile. Most managed NoSQL services commit to availability alone.^{46, 47}

• Search built into the engine: Vector search and the DiskANN index reached general availability in 2024, and full-text and hybrid search followed at Build 2025, so semantic retrieval runs in the same container as the operational data.^{48, 49}

Top features:

• Global distribution across any number of Azure regions⁴⁶

• Single-digit millisecond read and write latencies at the 99th percentile within a region⁴⁶

• Multi-model support with APIs for NoSQL, MongoDB, Cassandra, Gremlin, and Table⁴⁵

• Vector search with flat, quantized flat, and DiskANN index types, plus full-text and hybrid search using BM25 ranking and reciprocal rank fusion^{48, 49, 50}

• Autoscale and serverless throughput modes, with one-click conversion from serverless to provisioned without downtime or data migration⁴⁹

Benefits:

• Availability commitments scale with topology, from 99.99% for a single-region account to 99.999% for reads on any multi-region account, and 99.999% for writes once multi-region writes are enabled.⁴⁷

• Throughput scales automatically with usage, and the service is integrated into Azure identity, networking, and monitoring rather than deployed alongside them.^{46, 49}

• Wire-protocol compatibility means MongoDB and Cassandra applications point at Cosmos DB with driver changes rather than a rewrite, and the NoSQL API uses SQL-like syntax familiar to relational developers.^{45, 46}

Cons:

• Strong and bounded staleness halve read throughput. Microsoft's documentation states that these levels read from two replicas in a four-replica set, so for the same number of request units you get half the read throughput of session, consistent prefix, or eventual. 45

• Costs escalate if not managed appropriately. Billing is per request unit rather than per node, so expense tracks query complexity as well as volume, and each additional region multiplies the provisioned throughput you pay for.⁴⁶

• New capabilities land on the NoSQL API first. Vector, full-text, and hybrid search all shipped there, so teams on the MongoDB, Cassandra, Gremlin, or Table APIs get compatibility at the cost of feature parity.^{48, 49, 50}

Best for: Organizations that need a wide range of development options for applications such as chatbots, fintech, IoT, and e-commerce.

In the same way that Google Bigtable is good for organizations using Google and Amazon DynamoDB is good for organizations using AWS, Microsoft Azure Cosmos DB is good for organizations using Microsoft Azure. That said, its wide variety of data and consistency models, as well as its multiple query languages, give developers more familiarity and flexibility.

MongoDB Atlas

MongoDB Atlas is the managed cloud service for MongoDB, the document database that has been the default NoSQL answer for most of the last decade. Data is stored as BSON documents in collections, grouped into replica sets for availability and sharded horizontally for scale.⁵²

Why it stands out:

• Multi-cloud by default: Atlas runs on AWS, Google Cloud, and Microsoft Azure, and one cluster can place replicas across more than one provider, which is a stronger form of portability than deploying separately on each.⁵²

• Deep ecosystem: Drivers, ODMs, and framework integrations exist for essentially every major language and stack, and the pool of developers who already know the query language is the largest of any database on this list.⁵²

• Owns its embedding models: MongoDB acquired Voyage AI in February 2025 and has been folding its embedding and reranking models into Atlas. No other vendor on this list owns the models that generate the vectors it stores.⁵⁷

Top features:

• Document data model⁵²

• Multi-document ACID transactions spanning multiple operations, collections, databases, documents, and shards⁵³

• Atlas Search for full-text and Atlas Vector Search for semantic retrieval, both running against the same cluster as operational data^{52, 56}

• MongoDB Query Language and the aggregation pipeline for multi-stage transformation and analysis in the database⁵²

• Integrated security including role-based access control, network isolation, and encryption at rest and in transit⁵²

Benefits:

• The document data model is built for flexibility and speed of development, because the structure of the data follows the application rather than the reverse.⁵²

• The size of the ecosystem means tooling, tutorials, and experienced developers are easy to find, though the server itself is source-available under the SSPL rather than open source.^{52, 54}

• Replica sets provide automatic failover, so the loss of a primary node is handled by the cluster rather than by staff. ⁵²

Cons:

• MongoDB is source-available, not open source. Community Server has been licensed under the SSPL since October 16, 2018, and MongoDB's own FAQ confirms the license is not OSI-approved, which rules out offering it as a managed service without a commercial agreement⁵⁴

• Write scaling depends on shard key choice. MongoDB's documentation warns that a suboptimal shard key produces jumbo chunks, uneven load distribution, and query performance that degrades over time. Resharding has been possible since version 5.0, but it is a heavy operation, so the decision carries weight up front⁵⁵

• The recommended pattern pushes complexity into schema design. MongoDB's documentation notes that single-document atomicity means you don’t need multi-document transactions for many practical use cases. However, that means embedding related data rather than normalizing it, and getting that structure wrong is expensive to undo later⁵³

Best for: Teams looking for a NoSQL database with data structure flexibility, especially those operating in a multi-cloud environment or needing a rich feature set to support diverse application requirements.

There are advantages to using something popular. Tooling exists, hiring is easier, and someone has already encountered whatever problem you are about to have. But popularity is not a fit assessment. What matters here is whether your write patterns suit a single primary per shard, and whether you can commit to a schema design early, because both are harder to change later.

Oracle NoSQL Database Cloud Service

Oracle is synonymous with SQL, but it offers a fully managed NoSQL service as well. It handles document, key-value, and fixed-schema data with single-digit millisecond response times, ACID transactions, and active-active regional replication, all on Oracle Cloud Infrastructure.⁵⁸ Pricing is pay-per-use across both on-demand and provisioned capacity modes.⁵⁸

Why it stands out:

• On-premises and cloud parity: Oracle advertises 100% compatibility between the cloud service and on-premises Oracle NoSQL Database, reachable through one programmatic interface with no application code changes, so a hybrid or repatriation move is not a rewrite.^{58, 60}

• Global Active Tables: Tables can be replicated across regions with local reads and writes in each, and Oracle provides built-in conflict resolution when the same record is updated in two regions at once.⁶⁰

• Oracle-native security and identity: Authorization and data access run through Oracle Cloud Identity and Access Management rather than a separate credential system, which matters if the rest of your estate is already governed there.⁵⁸

Top features:

• Multi-model support for document, key-value, and fixed-schema data^{58, 60}

• Adjustable durability and consistency per operation, with absolute and eventual consistency both available^{60, 61}

• SQL queries over schema-less JSON, with secondary indexes on any supported field and server-side atomic updates to parts of a JSON document⁵⁹

• On-demand and provisioned capacity modes, so throughput and storage can be scaled up during peaks and back down afterward⁵⁹

• Time-to-live on table rows, so records expire automatically without a cleanup job⁵⁹

Benefits:

• Oracle manages the infrastructure, software, security, capacity scaling, operations, and maintenance, including upgrades, patches, and hardware failures.^{59, 60}

• Consistency is tunable per operation, so latency-sensitive reads can be relaxed to eventual consistency while transactional paths keep ACID guarantees.⁶⁰

• Predictable single-digit millisecond response times, with data replicated for high availability.⁵⁸

Cons:

• Absolute consistency costs twice as much as eventual. Oracle's capacity documentation shows a 1.5 KB record using two read units for an eventually consistent read and four for an absolutely consistent one⁶¹

• Vector search is not in this service. Oracle's vector capability lives in Oracle AI Database, so semantic retrieval alongside NoSQL data means running and syncing two Oracle products rather than one

• The Always Free tier is available in one region. Oracle's documentation restricts Always Free NoSQL Database Service to Phoenix, which makes low-cost evaluation awkward for teams outside that region⁵⁹

Best for: Enterprises that already rely on a range of Oracle products.

Organizations that already use Oracle relational databases and other products might find Oracle NoSQL Database Cloud Service an easier sell if they are looking to experiment with NoSQL. Its multiple data models, consistency policies, and SQL-like query language act as a familiar on-ramp.

Redis Enterprise Cloud

Initially created as an in-memory cache store to boost performance, Redis has evolved over the years. Today it serves as a high-speed, low-latency NoSQL database with search, JSON, time series, and vector types built into the core engine.⁶⁴ Its licensing has changed several times since 2024: Redis left the BSD license in March 2024 for the source-available RSALv2 and SSPLv1, then added AGPLv3 as a third option with Redis 8 in May 2025.⁶³

Why it stands out:

• Speed: Its in-memory architecture makes it fast, particularly for data that's frequently accessed. Redis positions predictable performance and linear scalability as core properties of the managed service.⁶²

• Local writes in every region: Active-active databases accept reads and writes in each geographic copy with sub-millisecond local latency, and conflicts between concurrent writes are resolved automatically by conflict-free replicated data types rather than by your application.^{66, 67}

• One engine, many data structures: Redis 8 merged Redis Stack into the core product, so search, JSON, time series, vector, and five probabilistic structures ship in the database itself rather than as modules you add.^{64, 65}

Top features:

• In-memory data store for sub-millisecond data serving^{62, 71}

• Active-active geo-distribution with multi-primary replication and strong eventual consistency across regions^{66, 67}

• Redis Flex, which tiers data across RAM and SSD so datasets can exceed available memory at lower cost per gigabyte⁶⁹

• A wide array of native data structures, including strings, hashes, lists, sets, sorted sets, and streams, plus JSON, time series, and probabilistic types such as Bloom filter, cuckoo filter, top-k, count-min sketch, and t-digest^{64, 71}

• Redis Query Engine for full-text, numeric, and vector queries, plus the vector set data type for lightweight similarity search, and hybrid search via FT.HYBRID as of 8.4^{65, 68}

Benefits:

• Serving the hot working set from memory removes the disk round trip that dominates latency in most database reads.^{62, 71}

• Available across major public cloud providers, with the same engine deployable on-premises as Redis Software, which suits hybrid organizations.^{62, 70}

• Instant failover, backups, and recovery are handled by the service, with 24/7 monitoring and support.⁶²

Cons:

• Licensing is complicated. Redis left BSD in March 2024 for RSALv2 and SSPLv1, then added AGPLv3 as a third option with Redis 8 in May 2025. The change prompted the Valkey fork, which some organizations are now evaluating instead.^{63, 64}

• Memory is the expensive resource, so cost tracks working-set size rather than request volume. Redis Flex mitigates this by tiering to SSD, but it adds a storage tier to size and manage.⁶⁹

• Active-active carries limits. Redis Cloud caps Active-active subscriptions at 10 regions and 10 databases, and the consistency model is strong eventual consistency, so CRDT conflict resolution is predictable but may not match your application's intended semantics.^{66, 67}

Best for: Frequently accessed data where latency is the binding constraint, particularly caching, session state, real-time leaderboards, and semantic caching for AI applications.

Organizations that rely on real-time analytics or need to manage live sessions with users find that Redis Enterprise Cloud performs well.

ScyllaDB X Cloud

ScyllaDB is a NoSQL database designed for data-intensive applications that require high throughput, low latency, and predictable scalability. It supports wide-column and key-value modeling from hundreds of gigabytes to petabytes with high availability. Its licensing changed in December 2024, when ScyllaDB moved to a source-available license and made OSS AGPL 6.2 its final open-source release.⁷⁶

Why it stands out:

• Shard-per-core performance: ScyllaDB assigns one shard per CPU core and reports steady single-digit millisecond P99 latency at workloads exceeding two million operations per second.^{72, 79}

• Tablets-based elasticity: Data is split into tablets that Raft rebalances dynamically, so clusters scale in parallel rather than serially. ScyllaDB reports scaling from 100,000 to two million operations per second in minutes and running safely at up to 90% storage utilization.^{72, 74, 75}

• Proven in real-time machine learning serving: Tripadvisor runs ScyllaDB on AWS for personalization, handling roughly 500,000 operations per second at peak with P99 latency of 1 to 3 milliseconds, and serving up to 5 million static features and 0.5 million user features per second from its feature store.^{80, 82}

Top features:

• Compatible with Apache Cassandra, including support for Cassandra Query Language, offering a transition for existing applications⁷²

• Advanced caching mechanisms improve read and write speeds⁷²

• Automatic scaling that adjusts node count and instance size as storage demand rises and falls, with no manual capacity planning or cluster rebalancing⁷³

• Support for key-value, wide column, and time series data formats⁷²

• Vector search on dedicated nodes with a RAM-resident index, decoupled from the operational storage engine so heavy vector queries cannot affect database latency, supporting embeddings up to 16,000 dimensions and tuned ANN indexes at billion-vector scale^{80, 81}

Benefits:

• Latency stays predictable as throughput rises, which matters when an SLA is measured at P99 rather than on average.⁷²

• Non-hierarchical architecture makes scaling simple, because all nodes are equivalent and replication and repartitioning are automatic.⁷⁸

• It integrates well with existing Cassandra workloads for ease of migration.⁷²

Cons:

• The data model is narrow. ScyllaDB covers wide-column, key-value, and time series, with no native document, JSON, or graph model, so it is a poorer fit than Couchbase, MongoDB, or Cosmos DB for schema-flexible application data.⁷²

• Licensing tightened in December 2024. ScyllaDB moved to a source-available license, with OSS AGPL 6.2 as the final open-source release, and the free tier of the full-featured Enterprise edition is capped at 10TB of total storage and 50 vCPUs per organization starting in 2025.^{76, 77}

• Scylla X Cloud does not yet support multi-datacenter deployments. ScyllaDB's documentation states multi-DC support is in progress, so the newest cluster type and multi-region topology are currently an either-or choice.⁷³

Best for: Companies with large data streams needing real-time processing, such as in financial services, AdTech, or IoT, which require low latency.

Organizations that anticipate rapid growth will find ScyllaDB X Cloud's auto-tuning, replication, and repartitioning optimizations to be critical assets that don't require as much operator time.

How to choose

Three questions help determine the right choice.

Where does your consistency requirement stop? DynamoDB will give you strong consistency across exactly three regions and no transactions with it. Couchbase gives you ACID inside a cluster and eventual consistency between them. Cosmos DB gives you five levels, but its documentation notes that the two strongest halve your read throughput per request unit. Astra DB is eventually consistent by design. If your application needs a guarantee that a specific write is visible everywhere before you act on it, that constraint eliminates most of this list.

What does your bill track? Request-based pricing, as in DynamoDB and Cosmos DB, means expense follows traffic, which is efficient for spiky workloads but hard to forecast. Capacity-based pricing, as in Aerospike Cloud and Bigtable, means expense follows provisioned footprint, which is predictable but wasteful at low utilization. Neither is better. They depend on how well you know your traffic.

How locked in are you willing to be? DynamoDB, Bigtable, Cosmos DB, and Oracle NoSQL are each excellent inside their own cloud and a re-platform outside it. Couchbase, MongoDB, Redis, ScyllaDB, Astra DB, and Aerospike run in more than one place. If that portability is worth something to you, it is worth pricing now.

Two other factors are also worth considering:

• Consolidation and license tightening have moved in one direction since 2024, and nothing on this list has moved back, so the licensing and ownership column deserves the same weight you would give a technical constraint.

• Vector search has arrived inside general-purpose operational databases rather than staying in specialist ones, which means the case for running a separate vector store is now narrower than it was when we first published this guide.

Whatever you choose, the decision gets easier once it stops being theoretical. The next step is picking two or three choices from the table above and testing them against your own access patterns, at the volumes you expect a year from now rather than the ones you have today. The differences that matter show up under load, and they show up faster than you might expect.