Standalone REST Proxy Deployment

This section is not meant to be an exhaustive guide to running the Confluent REST Proxy, but it covers the key things to consider before you should consider the proxy production ready.

Three main areas are covered:

  • Logistical considerations, such as hardware recommendations and deployment strategies
  • Configuration changes that are more suited to a production environment
  • Post-deployment considerations


If you’ve been following the normal development path, you’ve probably been playing with the REST Proxy on your laptop. But when it comes time to deploying to production, there are a few recommendations that you should consider.


The REST Proxy’s memory usage is primarily tied to the number of consumers because these are the only stateful resources managed by the proxy. The consumer buffers messages in two ways that can affect total memory usage. First, the underlying Java consumer buffers up to fetch.max.message.bytes x queued.max.message.chunks bytes of data, with default values resulting in 2 MB per consumer. Second, during each consumer request, up to consumer.request.max.bytes bytes may be buffered before the response is returned; the default value is 64 MB. In practice, the average memory usage per consumer is closer to the first value because most consumers will either have a steady stream of data, in which case requests return quickly instead of buffering up to consumer.request.max.bytes byte or they have little data coming through and therefore use little buffer space.

All produce requests are processed by a single set of producers, one per data format. Each has a buffer of records waiting to be sent, by default 32 MB each. With the current default producer settings and data formats (binary, Avro, JSON schema, Protobuf), this requires only 64 MB. If you are using Avro, JSON schema, or Protobuf, the serializer in the producer and deserializers in consumers also maintain a cache of schemas. However, schemas are relatively small and so should not significantly affect memory usage.

If you plan to use the REST Proxy mainly for administrative actions or producing data to Kafka, the memory requirements are modest, and a heap size of 1GB would suffice. If you plan to use many consumers, you can do a back of the envelope calculation to determine a reasonable heap size based on the maximum number of consumers you expect and average memory usage of ~16 MB per consumer when using the default configuration.


The CPU requirements for the REST Proxy mirror those of normal clients: the major computational costs come from compression and serialization of messages. The REST Proxy can process many requests concurrently and can take advantage of more cores if available. We recommend at least 16 cores, which provides sufficient resources to handle HTTP requests in parallel and background threads for the producers and consumers. However, this should be adjusted for your workload. Low throughput deployments may use fewer cores, while a proxy that runs many consumers should use more because each consumer has a dedicated thread.


The REST Proxy does not store any state on disk. The only disk usage comes from log4j logs.


A fast and reliable network will likely have the biggest impact on the REST Proxy’s performance. It should only be used as a proxy for Kafka clusters in the same data center to ensure low latency access to the Kafka brokers. Standard data center networking (1 GbE, 10 GbE) is sufficient for most applications.


You should use Java 17, although Java 11 and Java 8 are supported. Java 8 has been deprecated and will be removed in a future version. To learn more about Java support in Confluent Platform, see Confluent Platform Java requirements.

The recommended GC tuning looks like the following:

-Xms1g -Xmx1g -XX:MetaspaceSize=96m -XX:+UseG1GC -XX:MaxGCPauseMillis=20 \
       -XX:InitiatingHeapOccupancyPercent=35 -XX:G1HeapRegionSize=16M \
       -XX:MinMetaspaceFreeRatio=50 -XX:MaxMetaspaceFreeRatio=80

The heap size setting of 1 GB should be increased for proxies that will use many consumers. However, instead of heap sizes larger than 8 GB we recommend running multiple instances of the REST Proxy to avoid long GC pauses that can cause request timeout and consumer disconnections. Use the environment variable KAFKAREST_HEAP_OPTS to set the heap size.

Deployment and Load Balancing

The REST Proxy does not require any coordination between instances, so you can easily scale your deployment up or down. The only requirement for multiple instances is that you set a unique id for each instance.

If you run more than one instance of REST Proxy you should provide some load balancing mechanism. The simplest approaches use round-robin DNS or a discovery service to select one instance per application process at startup, sending all traffic to that instance. You can also use an HTTP load balancer, but individual instances must still be addressable to support the absolute URLs returned for use in consumer read and offset commit operations.

Generally, you need a sticky load balancer session because consumers are stateful. Therefore, out-of-the-box REST Proxy requires that the load balancer is sticky; that is, each consumer instance should always be routed to the same REST Proxy instance.

However, you can use a stateless connection if your consumer sends only the initial request to the load balancer, and then uses the returned hostname on subsequent requests.

High Availability

When deploying multiple REST Proxy instances it is important that your consumer client handle exceptions returned from a failed consumer instance. Following an exception your client should attempt to create a new consumer instance using the load balancer address in order to switch to an active REST Proxy instance. As noted above the client should use the absolute URL returned after creating a consumer instance for further communication. Producing messages through multiple REST Proxy instances doesn’t require this level of exception handling.

Your client should attempt to delete all consumer instance before exiting. This helps avoid zombie consumer instances that will cause consumption delay based on the value of in the REST Proxy configuration properties. The default is equivalent to 5 minutes (300,000ms) and depending on the stability of your client it may be worth lower this value.

Important Configuration Options

The full set of configuration options are documented here .

However, some configurations should be changed for production. Some must be changed because they depend on your cluster layout:


A list of host/port pairs to use for establishing the initial connection to the Kafka cluster. The client will make use of all servers irrespective of which servers are specified here for bootstrapping. This list only impacts the initial hosts used to discover the full set of servers. This list should be in the form host1:port1,host2:port2,.... Because these servers are only used for the initial connection to discover the full cluster membership (which may change dynamically), this list does not require the full set of servers. You might want to specify multiple servers in case one goes down.

  • Type: list
  • Default:
  • Valid Values:
  • Importance: high

The base URL for Schema Registry that should be used by the serializer.


The configuration property auto.register.schemas is not supported for Kafka REST Proxy.


Unique ID for this REST server instance. This is used in generating unique IDs for consumers that do not specify their ID. The ID is empty by default, which makes a single server setup easier to get up and running, but is not safe for multi-server deployments where automatic consumer IDs are used.

  • Type: string
  • Default: “”
  • Importance: high

Other settings are important to the health and performance of the proxy. You should consider changing these based on your specific use case.


Maximum number of bytes in message keys and values returned by a single request. Smaller values reduce the maximum memory used by a single consumer and may be helpful to clients that cannot perform a streaming decode of responses, limiting the maximum memory used to decode and process a single JSON payload.

Conversely, larger values may be more efficient because many messages can be batched into a single request, reducing the number of HTTP requests (and network round trips) required to consume the same set of messages.

Note that this can also be overridden by clients on a per-request basis using the max_bytes query parameter. However, this setting controls the absolute maximum; max_bytes settings exceeding this value will be ignored.

  • Type: long
  • Default: 67108864
  • Importance: medium

The minimum number of bytes in message keys and values returned by a single request before the timeout of passes.

  • Type: int
  • Default: -1
  • Importance: medium

The maximum total time to wait for messages for a request if the maximum request size has not yet been reached. The consumer uses a timeout to enable batching. A larger value will allow the consumer to wait longer, possibly including more messages in the response. However, this value is also a lower bound on the latency of consuming a message from Kafka. If consumers need low latency message delivery, then specify a lower value.

  • Type: int
  • Default: 1000
  • Importance: medium

The maximum number of threads to run consumer requests on. Consumers requests are ran one per thread in a synchronous manner. You must set this value higher than the maximum number of consumers in a single consumer group, otherwise rebalances will deadlock.

  • Type: int
  • Default: 50
  • Importance: medium

The host name used to generate absolute URLs for consumers. If empty, the default canonical hostname is used. You may need to set this value if the FQDN of your host cannot be automatically determined.

  • Type: string
  • Default: “”
  • Importance: medium

Don’t Touch These Settings!

Changing the following settings may lead to very poor performance. They have been selected carefully to balance important performance tradeoffs. If you do need to change them, test the configuration very thoroughly before putting it into production!

Amount of time to back off when an iterator runs out of data. If a consumer has a dedicated worker thread, this is effectively the maximum error value for the entire request timeout. It should be small enough to closely target the timeout, but large enough to avoid busy waiting.

  • Type: int
  • Default: 50
  • Importance: low

Timeout for blocking consumer iterator operations. This should be set to a small enough value that it is possible to effectively peek() on the iterator.

  • Type: int
  • Default: 1
  • Importance: low

Post Deployment

Although the proxy does not have any persistent state, it is stateful because consumer instances are associated with specific proxy instances. If a proxy process has consumers that are part of a consumer group, shutting down or restarting that proxy will cause a rebalance operation for the remaining consumers. This event is expected and isolated instances, for example due to a hardware failure or network outage, will not cause problems. However, operators should be aware that this rebalance is not instantaneous and needs to be accounted for in site-wide updates, such as rolling restarts of all REST proxies for updates.

Upgrades to newer versions are simple because there is no persistent state. A rolling restart of all servers, leaving sufficient time for rebalance operations as described above, is a safe way to perform a zero-downtime upgrade.