You are viewing an old version of this page. View the current version.

Compare with Current View Page History

« Previous Version 12 Next »

Status

Current State: Draft

Discussion Thread: link

JIRA: KAFKA-4011

Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).

Motivation

Kafka currently supports setting an upper bound on the number of requests allowed into the (incoming) request queue. This is an indirect way of controlling memory consumption and has a few drawbacks:

  1. An administrator needs to estimate the average request size in order to provide a meaningful size limit.
  2. This size limit may need to be periodically updated as the workload changes.
  3. The server is still susceptible to a simultaneous batch of large requests exhausting the JVM memory (causing an OOM exception).

The third scenario actually occurred a few times at LinkedIn - a sudden spike of very large request batches (1000s requests each) from a Hadoop job caused OOM exceptions on a production cluster.

This KIP proposes allowing an administrator to specify a memory limit in bytes, which resolves the above problems.

Public Interfaces

This KIP introduces a new server configuration parameter, queued.max.bytes, that would specify a limit on the volume of requests that can be held in memory. This configuration parameter will co-exist with the existing queued.max.requests (the code will respect both bounds and not picking up new requests when either is hit).

Beyond the proposed new configuration key this KIP makes no changes to client or server public APIs.

Proposed Changes

a new MemoryPool interface would be introduced into the kafka codebase:

MemoryPool.java
public interface MemoryPool {
    /**
     * Tries to acquire a ByteBuffer of the specified size
     * @param sizeBytes size required
     * @return a ByteBuffer (which later needs to be release()ed), or null if no memory available.
     *         the buffer will be of the exact size requested, even if backed by a larger chunk of memory
     */
    ByteBuffer tryAllocate(int sizeBytes);

    /**
     * Returns a previously allocated buffer to the pool.
     * @param previouslyAllocated a buffer previously returned from tryAllocate()
     */
    void release(ByteBuffer previouslyAllocated);
}
  1. the pool is non-blocking, so network threads would not be blocked waiting for memory and could make progress elsewhere.
  2. SocketServer would instantiate and hold a memory pool, which Processor threads would try to allocate memory from when reading requests out of sockets (by passing the pool to instances of NetworkReceive that they create).
  3. NetworkReceive.readFromReadableChannel() would be modified to try allocating memory (it is already written in a way that reading may involve multiple repeated calls to readFromReadableChannel(), so not a big change to behavior)
  4. memory would be released at the end of request processing (in KafkaRequestHandler.run())
  5. to facilitate faster implementation (as a safety net) the pool will be implemented in such a way that memory that was not release()ed (but still garbage collected) would be detected and "reclaimed". this is to prevent "leaks" in case of code paths that fail to release() properly.

Caveats

  1. As the pool would allow any size request if it has any capacity available, the actual memory bound is queued.max.bytes + socket.request.max.bytes. The up-side is no issues with large requests getting starved out
  2. concerns have been raised about the possibility of starvation in Selector.pollSelectionKeys() - in case the order of keys in Set<SelectionKey> selectionKeys is deterministic and memory is tight, sockets consistently at the beginning of the set get better treatment then those at the end of the iteration order. to overcome this code has been put in place to shuffle the selection keys and handle them in random order ONLY IF MEMORY IS TIGHT (so if the previous allocation call failed). this avoids the overhead of the shuffle when memory is not an issue.

Compatibility, Deprecation, and Migration Plan

There are a few approaches w.r.t migration. The current preference is to go with the third option.

  1. queued.max.requests is deprecated/removed in favor of queued.max.bytes. In this case, the conversion of existing configurations could use queued.max.bytes = queued.max.requests * socket.request.max.bytes (which is conservative, but "safe")
  2. queued.max.requests is supported as an alternative to queued.max.bytes (either-or), in which case no migration is required. A default value of 0 could be used to disable the feature (by default) and runtime code would pick a queue implementation depending on which configuration parameter is provided.
  3. queued.max.requests is supported in addition queued.max.bytes (both respected at the same time). In this case a default value of queued.max.bytes = -1 would maintain backwards compatible behavior.

The current naming scheme of queued.max.requests (and the proposed queued.max.bytes) may be a bit opaque. Perhaps using requestQueue.max.requests and requestQueue.max.bytes would more clearly convey the meaning to users (indicating that these settings deal with the request queue specifically, and not some other). The current queued.max.requests configuration can be retained for a few more releases for backwards compatibility.

Configuration Validation

queued.max.bytes must be larger than socket.request.max.bytes (in other words, memory pool must be large enough to accommodate the largest single request possible), or negative (if disabled).

Test Plan

  • A unit test was written to validate the behavior of the memory pool
  • A unit test that validates correct behavior of RequestChannel under capacity bounds would need to be written.
  • A micro-benchmark for determining the performance of the pool would need to be written
  • Stress testing a broker (heavy producer load of varying request sizes) to verify that the memory limit is honored.
  • Benchmarking producer and consumer throughput before/after the change to prove that ingress/egress performance remains acceptable.

Rejected Alternatives

Here are some alternatives that we have discussed (at LinkedIn):

  1. Reducing producer max batch size: this is harmful to throughput (and is also more complicated to maintain from an administrator's standpoint than simply sizing the broker itself). This is more of a workaround than a fix
  2. Reducing producer max request size: same issues as above.
  3. Limiting the number of connected clients: same issues as above
  4. Reducing queued.max.requests in the broker: Although this will conservatively size the queue it can be detrimental to throughput in the average case.
  5. controlling the volume of requests enqueued in RequestChannel.requestQueue (would not suffice as no bound on memory read from actual sockets)

Implementation Concerns

  1. the order of selection keys returned from a selector.poll call is undefined. in case the actual implementation uses a fixed order (say by increasing handle id?) and under prolonged memory pressure (so never enough memory to service all requests) this may lead to starvation of sockets that are always at the end of the iteration order. to overcome this the code shuffles the selection keys if memory is low.
  2. a strict pool (which adheres to its max size completely) will cause starvation of large requests under memory pressure (as they would never be able to allocate if there is a stream of small requests). to avoid this the pool implementation will allocate the requested amount of memory if it has any memory available (so if pool has 1 free byte and 1 MB is requested, 1MB will be returned and the number of available bytes in the pool will be negative). this means the actual bound on number of bytes outstanding is queued.max.bytes + socket.request.max.bytes - 1 (socket.request.max.bytes representing the largest single request possible)
  3. if there is no memory available at all in Selector.poll() the select() call will return immediately with read-ready channels that we cannot service (because no memory available). this may cause the SocketServer.Processor.run() call (which calls into poll) to go into a tight loop where no progress can be made. it is my (Radai) opinion that in such a scenario there will very likely be progress to be made elsewhere in run() (for example processCompletedSends() - presumably memory is tight because a lot of requests are executing, some of them are very likely done). to avoid the tight loop code has been written to mute all channels when no memory is available (so the select() call will block for a while waiting for other things) and unmute them if/when memory becomes available in a future call to poll(). this code is available in a separate branch. issues with this implementation:
    1. when memory is unavailable channels will be muted until the next call to poll(). if no other channel activity is present except reads this means a 300ms wait at least (thats the current time a poll() call is set to wait)
    2. every such mute/unmute cycle (in response to memory becoming unavailable/available accordingly) is O(#channels), which seems excessive.
    perhaps a better alternative would be to have reads (as opposed to connects, disconnects and writes) be done by dedicated threads. such threads could actually block waiting for memory (something that currently SocketServer.Processor cannot do as it has more responsibilities beyond reads)

State of Implementation

an implementation is available - https://github.com/radai-rosenblatt/kafka/tree/broker-memory-pool

a separate branch implements channel muting/unmuting under memory pressure - https://github.com/radai-rosenblatt/kafka/tree/broker-memory-pool-with-muting

 

  • No labels