Ververica

Editor’s note: This Ververica review is based on my hands-on evaluation of Ververica Cloud BYOC (release 10.2026, VERA 4.6) and Self-Managed Platform 3.1.4 in October 2026, weighing pricing, setup, VERA, connectors, AI SQL, operations and rivals.

Quick verdict: I recommend Ververica for regulated Flink shops that need Flink in their own cloud account or data center, backed by a 99.99% BYOC SLA. The trade-off: no managed option for new customers, BYOC bills $0.28 per CU-hour on top of your cluster, and Self-Managed is quote-only.

Flink’s creators founded Ververica in Berlin in 2014 as data Artisans; Alibaba bought it in January 2019, and it is now based in Munich. Its VERA engine, Ververica says, runs existing Flink JARs unchanged. In this review, I’ll take a closer look at Ververica’s pricing, features, and deployment options, so you can see exactly whether that support is worth the bill.

Key Takeaways:

• Two models: BYOC in your EKS or AKS cluster, or Self-Managed on any Kubernetes, air-gapped included.
• BYOC on AWS: $0.28 per CU-hour (1 CPU core, 4 GiB) plus AWS infrastructure.
• No managed entry: the multi-tenant service closed to new users on June 30, 2026; no free trial.
• Version lag: VERA 4.6 runs on Flink 1.20, while Apache Flink reached 2.3 in June 2026.
• Fast releases: BYOC 10.2026 shipped October 22, 2026.

Pros and Cons

How Much Does Ververica Cost?

Ververica no longer publishes a price list on its own site, so the only public rate is its BYOC subscription on AWS Marketplace: $0.0028 per centiCU-hour, or $0.28 per CU-hour. One CU is 1 CPU core and 4 GiB of memory, billed hourly on CUs consumed with no commitment and no contract tiers. AWS bills your EKS, EC2 and storage separately.

• BYOC on AWS ($0.28 per CU-hour): for teams that keep the data plane in their own EKS cluster. AWS bills the usage with no separate Ververica invoice, Marketplace usage gets no free credits or promotions, and you can cancel any time.
• BYOC on Azure (Azure Marketplace plans): for AKS teams, live since February 18, 2026. Plans run pay-as-you-go, monthly, annual or multi-year, with consolidated billing and drawdown of Azure consumption commitments.
• Self-Managed Ververica Platform (custom quote): for any Kubernetes, on-premises included. The license goes through sales, with three support plans:
• Bronze: business hours Monday to Friday, 9 to 18, with P1 response within 8 business hours.
• Silver: P1 and P2 response within 4 business hours.
• Gold: 24/7 production support, with P1 and P2 response within 1 hour.

The multi-tenant Managed Service closed to new users on June 30, 2026. Existing customers keep access, and Ververica says serverless will return, with no date attached.

Option	Price	Billing	Where it runs	Support/SLA
BYOC on AWS	$0.28 per CU-hour plus AWS infrastructure	Hourly usage through AWS Marketplace	Your EKS cluster	99.99% platform SLA
BYOC on Azure	Azure Marketplace plans	Pay-as-you-go or commitment plans	Your AKS cluster	99.99% platform SLA
Self-Managed	Custom quote	Negotiated license	Any Kubernetes	Bronze, Silver or Gold plan

Amazon Managed Service for Apache Flink charges $0.11 per KPU-hour in US East, and a KPU is 1 vCPU and 4 GB, the same size as a Ververica CU. Amazon’s rate includes the compute, while Ververica’s $0.28 sits on top of EC2 nodes you pay for. Amazon adds one orchestration KPU per application and $0.10 per GB-month of running application storage.

Confluent Cloud for Apache Flink bills $0.21 per CFU-hour, metered per minute. A CFU is a logical unit rather than a fixed core count, so I compare it with Ververica on a real workload, not on the hourly rate.

Is Ververica Good Value for Money?

• Good fit: teams that must keep data inside their own account or data center and already run EKS, AKS or OpenShift.
• Less convincing fit: small teams that just want serverless Flink. Amazon charges less per unit and runs the cluster for you.
• Budget priority: model CU-hours plus the EKS node cost before committing. Autopilot can cut parallelism (see Operations below), but the node bill stays yours.

Reviewer’s Notes: Start BYOC pay-as-you-go on AWS Marketplace with one non-critical job to measure real CU-hours, then negotiate a commitment or a Self-Managed license from that usage.

My Experience With Ververica

A first running job on Ververica now starts with installing an agent into your own Kubernetes cluster. Confluent Cloud for Apache Flink and Amazon’s managed service have no cluster to install, so this is the step that separates Ververica from both.

Setting Up BYOC on AWS

I set up BYOC on Amazon EKS with EC2 nodes, since Fargate is not supported. The cluster had to meet these prerequisites:

• Kubernetes 1.28 or later on Linux nodes from families such as m5 or m6.
• Storage: gp2 or gp3, node root volumes of at least 100 GiB, and two 5 GB persistent volume claims per workspace.
• AWS resources: IAM roles and policies, an S3 bucket for Flink artifacts, and IRSA through an OIDC provider.
• Network: outbound HTTPS only, to agent.ververica.cloud, app.ververica.cloud, registry.ververica.cloud and cdn.ververica.cloud.

In the Ververica Dashboard profile menu, I went to Payment, connected AWS, subscribed to Ververica Cloud: Bring Your Own Cloud (PAYG) on AWS Marketplace and linked the accounts.

Next I registered the agent, downloaded the pre-filled values.yaml, added pyxisRoleArn under global and ran the Helm install. The registration token expires after 30 minutes, so I had the IAM roles and S3 bucket ready before I registered. With the agent verified, I created a BYOC workspace.

The Ververica Agent opens an outbound HTTPS tunnel to the control plane. Ververica never initiates access to the cluster, and its control plane stores only metadata.

Running My First Flink SQL Job

I used a datagen source and a print sink for the first job, which kept Kafka credentials out of day one:

1. On the Dashboard, I opened the workspace Console, then SQL Editor > New > Blank Stream Draft > Next.
2. I named the draft, kept the default Location and Engine Version, and clicked Create.
3. I pasted a datagen source table, a print sink with 'logger' = 'true' and an INSERT INTO statement.
4. I clicked Validate, then Deploy and Confirm.
5. Under Deployments, I clicked Start, chose Initial mode and started the job.

The status moved from STARTING to RUNNING, and the output appeared under Diagnostics > Stdout and Logs > Running Task Managers. I stopped the job from Deployments.

Where Setup Gets Fiddly

• Kafka keystores: files uploaded as artifacts land in /flink/usrlib/, so the path must read /flink/usrlib/kafka.client.keystore.jks with no file:// prefix, or the job throws NoSuchFileException.
• Artifact uploads: BYOC 10.2026 moved upload signing to the agent, so the S3 bucket CORS must allow the console origin, PUT and Content-Type.
• JAR size: Platform 2.15.12 raised connector and UDF uploads to 500 MB, but session clusters still reject JARs over 100 MiB.

Reviewer’s Notes: Have a Kubernetes admin on the call for the BYOC install. IAM, storage and the 30-minute agent window carry the setup work; the SQL side is five steps once the workspace exists.

The VERA Engine and Flink Compatibility

VERA is binary-compatible with Apache Flink and “not a fork”: Ververica says existing JARs, connectors and UDFs run unchanged and open-source Flink state snapshots carry over. Every VERA release maps to one Flink version, so the engine you pick must match your application’s Flink line. New deployments default to vera-4.6-jdk11-flink-1.20, and Java 11 and 17 variants exist for VERA 4.5 and 4.6.

VERA’s proprietary execution layer adds adaptive query planning, optimized memory management, operator fusion, an advanced state backend and JIT compilation. Ververica’s own figures are 2x throughput versus open-source Flink, 40% lower TCO and sub-10ms p99 latency. No public methodology accompanies that headline table, so I treat those numbers as targets to test on my own workload.

VERA-X is a separate vectorized engine with a ForStDB state store, offered to selected customers. Ververica claims 5 to 10 times the throughput of Flink 1.19 on Nexmark SQL benchmarks and 52% lower average resource use, but VERA-X does not yet support disaggregated state storage.

VERA’s Flink base trails upstream by a major version:

• VERA 4.x: built on Flink 1.20; the older VERA 1.0.x line sits on Flink 1.17.
• Apache Flink 2.3.0: released in June 2026, and Amazon’s managed service supported it from July 20, 2026.
• Self-Managed Platform 2.15.12: supports Flink 1.18, 1.19 and 1.20.

VERA itself ships often: 4.1 on November 7, 2025, 4.3 on February 9, 2026, 4.5 on May 11 and 4.6 on September 30. None of those releases moved VERA to Flink 2.x.

Upgrades within the line carry risk too. VERA 4.6 changed the state format for STDDEV and VAR aggregates, so jobs using them need a stateless restart unless table.exec.legacy-stddev-behaviour is set to true. Check your aggregates before you move a stateful job to 4.6.

Development route	Ververica deployment option
Flink SQL	SQL drafts and deployments
Java DataStream or Table API	JAR deployments
Python	Python deployments

Each route has its own quickstart. Open-source Apache Flink gives you 2.x features first, but no vendor engine and no support contract.

• Best for: established Flink 1.20 jobs that need a supported, compatible engine and someone to call.
• Skip if: your roadmap depends on Flink 2.x features today.

Connectors, CDC and the Streamhouse

Ververica markets 100+ connectors with Flink CDC, while its own connector catalog names about 22:

• Event streams: Apache Kafka, Upsert Kafka, Amazon Kinesis and Azure Event Hubs.
• Operational databases: MySQL, PostgreSQL and Oracle (each with a CDC source), SQL Server with Kerberos, MongoDB and Azure Cosmos DB.
• Serving and analytics: Elasticsearch, Redis, Snowflake, BigQuery, StarRocks and Milvus.
• Lakehouse storage: Apache Iceberg, Apache Paimon and Delta Lake.
• Test data: DataGen and Faker.

Catalogs cover Iceberg, Paimon, Delta and Fluss on the lake side, Kafka JSON and Kafka Schema Registry for streams, and MongoDB, MySQL and Oracle for databases.

Any open-source Flink connector outside that list can be uploaded as a JAR, and Platform 2.15.12 raised the connector and UDF upload limit to 500 MB. Some connectors exist only on certain deployment models, so confirm the connector, catalog and engine combination for yours before you plan a migration.

CDC Pipelines and Streamhouse Availability

Ververica donated Flink CDC to the Apache Software Foundation in March 2024. VERA 4.6 made YAML-defined CDC ingestion generally available for MySQL into Paimon, StarRocks and Kafka, with schema evolution, regex table routing and Kafka field routing, and the same release added the Azure Event Hubs and BigQuery connectors. MongoDB, PostgreSQL and Fluss in the same pipeline format are still public preview.

Streamhouse is Ververica’s name for Flink plus Apache Fluss streaming storage plus Paimon or Iceberg for long-term lake storage, with materialized tables available in SQL. Fluss became an Apache top-level project on July 16, 2026, but on Ververica it is Private Preview: free for subscribed customers and outside the SLA. Keep it out of SLA-bound pipelines for now.

Confluent‘s Flink is tightest with Kafka topics and Schema Registry, but it is regional with no cross-region queries. Ververica fits mixed estates of Kafka, CDC databases and lakehouse tables; Confluent fits pipelines that start and end in Confluent Kafka.

AI SQL Functions

Since VERA 4.5, Flink SQL on Ververica can call LLMs and ML models in-stream:

• Classification: AI_CLASSIFY and AI_SENTIMENT.
• Text processing: AI_EXTRACT, AI_SUMMARIZE, AI_TRANSLATE and AI_MASK.
• Embeddings and search: AI_EMBED and VECTOR_SEARCH.
• Model inference: ML_PREDICT, with model DDL such as CREATE TEMPORARY MODEL.

The functions run through LATERAL TABLE syntax and cannot be used as scalar expressions. Each needs a model created with CREATE MODEL that points at an external endpoint you configure, so latency depends on the model provider’s API, not on Flink.

VERA 4.6 (September 30, 2026) added functions that prepare images and PDFs for vision-language models such as Qwen-VL:

• Content handling: FETCH_CONTENT and MIME_TYPE.
• Image processing: IMAGE_COMPRESS and IMAGE_LAPLACE_VAR.
• PDF processing: PDF_TO_IMAGES.

The same release added the VARIANT type, and the Milvus connector handles vector writes. Two pipelines fit well: AI_MASK masking sensitive fields in events before they reach a sink, and AI_CLASSIFY or AI_EXTRACT enriching CDC records before they land in Paimon or Iceberg.

Amazon’s managed service leaves model calls to your own application code. I recommend Ververica’s AI functions for teams already writing Flink SQL who want the model call inside the pipeline definition instead of a separate service. Run them against a sample stream first to check the endpoint’s latency and output quality. Skip them if your models already run in-process inside your Flink jobs, because every one of these calls leaves the cluster for an external endpoint.

Operations, Security and Self-Managed Editions

Autopilot ships DISABLED by default and, once switched on per deployment, adjusts parallelism:

• DISABLED: no scaling.
• MONITORING: recommendations only.
• ACTIVE: applies recommendations, which upgrades the deployment with its configured upgrade strategy.

I recommend running MONITORING first, because every applied recommendation is a deployment upgrade. Savepoints and the stop options sit on the Deployments page.

Access control covers SAML and OIDC SSO, organizations, and workspace roles of Owner, Admin, Editor and Viewer. Platform 3.1.4 (October 13, 2026) added custom namespace roles, engine artifacts from your own blob storage and credentials from Kubernetes Secrets. BYOC 10.2026 (October 22, 2026) added an API that revokes a single SSO user’s sessions and tokens within about 10 seconds; revocation is API-only, not in the console.

Ververica holds ISO 27001:2022, SOC 2 Type II and SOC 3, and states a 99.99% platform SLA for BYOC, with preview features excluded. BYOC adds customer-managed KMS, BYOK encryption, TLS 1.3 and single-tenant deployment, and Ververica positions it for GDPR, DORA, NIS2, PCI DSS and HIPAA requirements.

Self-Managed Platform 2.x vs 3.x

Platform 2.15.12 (October 7, 2026) is the latest 2.x release, with Stream and Spring editions; Autopilot is only in Stream Edition and above. Platform 3.1.4 installs with Helm. Both lines run on EKS, AKS, GKE, OpenShift, Rancher RKE2 or vanilla Kubernetes, with an operator pattern, RBAC, SSO, backup and DR, offline licenses and air-gapped installs.

Area	Platform 2.x	Platform 3.x
Current release	2.15.12	3.1.4
Notable distinction	Stream and Spring editions	Custom namespace roles
Deployment control	Self-managed Kubernetes	Self-managed Kubernetes via Helm
Relevant limit or change	Autopilot in Stream Edition and above	Own-storage engine artifacts and Kubernetes Secrets

Platform 2.14.2 upgrades could lose checkpoint progress and hit a JobManager restart loop in HA setups, and jobs can fail to start on catalog connection timeouts. Put upgrade rollback in your acceptance plan.

Amazon’s managed service handles AZ failover and backups for you, on AWS only.

• Best for: Kubernetes-capable teams that need supported Flink operations, granular roles or air-gapped installs.
• Skip if: you want the vendor, not your platform team, to own cluster recovery.

How Does Ververica Compare to Competitors?

Ververica wins when Flink must run in your own account with vendor support; Amazon wins on unit price on AWS, Confluent on Kafka integration and RisingWave on Postgres-style SQL.

Tool	Starting price	Model	Best at
Ververica	BYOC $0.28 per CU-hour on AWS Marketplace plus infrastructure; Self-Managed quote	Flink-first platform in your account	Regulated, self-hosted Flink
Amazon Managed Service for Apache Flink	$0.11 per KPU-hour in US East, plus orchestration KPU	Serverless managed Flink on AWS	AWS-native teams, Flink 2.3
Confluent Cloud for Apache Flink	$0.21 per CFU-hour, metered per minute	Serverless Flink inside Confluent	Kafka-centric SQL
Open-source Apache Flink	$0 plus your ops	Framework	Full control, newest versions
RisingWave Cloud	From $0.227 per RWU-hour, 7-day trial	Streaming database	Postgres-compatible streaming SQL

A CFU or RWU is not a Ververica CU, and Ververica’s rate excludes the EC2 nodes, so compare on a real workload, not on the hourly column.

• Ververica vs Amazon Managed Service for Apache Flink: Amazon handles provisioning, automatic scaling, AZ failover and backups, supports Java, Scala, Python and SQL with Studio notebooks, and runs Flink 2.3 against VERA’s 1.20 base. Budget for its orchestration KPU, storage and durable backups, and two extra KPUs per Studio notebook. Ververica wins when the cluster must be yours or must leave AWS.
• Ververica vs Confluent Cloud for Apache Flink: Confluent wins when Kafka topics and Schema Registry already define the workload, with SQL and the Java Table API GA and the Python Table API in preview. Every Flink table there also creates a Kafka topic and schema, and a Baseline CFU is billed even when a statement sits idle. It is regional with no cross-region queries; Ververica takes existing Flink JARs and customer-run infrastructure.
• Ververica vs open-source Apache Flink: open-source Flink costs nothing in licenses and ships 2.x first, but cluster operations and recovery are yours. Pay for Ververica only when VERA, Autopilot and a support plan answer a concrete requirement.
• Ververica vs RisingWave: RisingWave suits Postgres-style streaming SQL when you don’t need Flink. Its Cloud Basic plan starts at $0.227 per RWU-hour and caps at 64 cores, its core is Apache 2.0, and self-managed RisingWave needs an annual contract. It cannot take Flink DataStream jobs or connector JARs, so it is a rewrite, not a migration.

Aiven for Apache Flink is closed to new services, and Redis agreed to buy Decodable in September 2025, which narrows the independent managed-Flink field.

How I Reviewed Ververica

I set up BYOC on AWS end to end: the EKS prerequisites, the Marketplace subscription and the agent install. In the console I created a SQL draft, then deployed, started and stopped the job and read its logs. I also worked through VERA and its Flink version line, connectors, CDC YAML pipelines and Streamhouse, the AI SQL functions, Autopilot, access control and compliance, the Self-Managed 2.x and 3.x editions, and every pricing route Ververica sells today.

I compared Ververica with Amazon Managed Service for Apache Flink, Confluent Cloud for Apache Flink, open-source Flink and RisingWave. Ververica’s throughput and TCO figures are its own, so I weighted deployment ownership and Flink version above them.

Prices current as of October 2026.

Should You Use Ververica for Apache Flink?

I recommend Ververica for supported Flink inside your own environment, particularly when your jobs fit VERA’s Flink 1.20 base. Its Kubernetes requirement and separate infrastructure bill make it a poor pick for teams that want cheap serverless Flink.

• Regulated teams (DORA, NIS2, data residency) on EKS, AKS or OpenShift: choose BYOC or Self-Managed, and validate your own controls rather than treating Ververica’s certifications as your compliance.
• Air-gapped or on-premises teams: choose Self-Managed with a Silver or Gold plan matched to your response-time needs.
• AWS-only teams wanting serverless Flink 2.3: choose Amazon’s managed service.
• Kafka-centric teams on Confluent: choose Confluent Cloud for Apache Flink.
• SQL-first teams that don’t need Flink: choose RisingWave.

For a team that already runs Kubernetes, measure before you commit, because the only public rate excludes your node bill. Subscribe to Ververica‘s BYOC pay-as-you-go plan on AWS Marketplace and port one existing Flink job to measure CU-hours before you talk to sales.

FAQs

Does Ververica have a free trial?

No. Ververica has no free trial for new customers, because the fully managed service and its trial closed to new users on June 30, 2026. BYOC on AWS Marketplace is pay-as-you-go with no commitment, but Marketplace usage gets no free credits, so budget for both Ververica’s CU-hours and the AWS infrastructure while you evaluate.

How much does Ververica cost?

Ververica BYOC on AWS Marketplace costs $0.28 per CU-hour ($0.0028 per centiCU-hour), plus AWS infrastructure. One CU is 1 CPU core and 4 GiB of memory. Azure runs through Azure Marketplace plans, and Self-Managed is a custom quote with Bronze, Silver or Gold support. See How Much Does Ververica Cost? above for rival rates.

Is VERA a fork of Apache Flink?

No. VERA is binary-compatible with Apache Flink, not a fork, and Ververica says existing JARs, connectors and UDFs run unchanged with compatible state snapshots. The VERA 4.x line is built on Flink 1.20, so features introduced in Flink 2.x are not available on it yet.

Which clouds does Ververica BYOC support?

Ververica BYOC runs on AWS EKS and Azure AKS, with Google Cloud listed as coming. Self-Managed is separate and runs on any Kubernetes, including GKE, OpenShift and on-premises clusters, with air-gapped installs. If you need GKE today, choose Self-Managed rather than waiting for Google Cloud BYOC.

Who owns Ververica?

Alibaba owns Ververica, having bought the company in January 2019. It was founded in Berlin in 2014 as data Artisans by the original creators of Apache Flink and is now headquartered in Munich. Apache Flink itself stays an open-source Apache project; VERA is Ververica’s proprietary engine on top of it.