This page is meant as a template for writing a KIP. To create a KIP choose Tools->Copy on this page and modify with your content and replace the heading with the next KIP number and a description of your issue. Replace anything in italics with your own description.

Status

Current state: Under Discussion

Discussion thread: https://www.mail-archive.com/dev@kafka.apache.org/msg83115.html

JIRA: KAFKA-6254

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

Motivation

Apache Kafka Brokers make periodic FetchRequests to other brokers, in order to learn about updates to partitions they are following.  These periodic FetchRequests must enumerate all the partitions which the follower is interested in.  The responses also enumerate all the partitions, plus metadata (and potentially data) about each one.

The frequency at which the leader responds to the follower's fetch requests is controlled by several configuration tunables, including replica.fetch.wait.max.ms, replica.fetch.min.bytes, replica.fetch.max.bytes, and replica.fetch.response.max.bytes.  Broadly speaking, the system can be tuned for lower latency, by having more frequent, smaller fetch responses, or for reduced system load, having having fewer, larger fetch responses.

There are two major inefficiencies in the current FetchRequest paradigm.  The first one is that the set of partitions which the follower is interested in changes only rarely.  Yet each FetchRequest must enumerate the full set of partitions which the follower is interested in.  The second inefficiency is that even when nothing has changed in a partition since the previous FetchRequest, we must still send back metadata about that partition.

These inefficiencies are linearly proportional to the number of extant partitions in the system.  So, for example, imagine a Kafka installation with 100,000 partitions, most of which receive new messages only rarely.  The brokers in that system will still send back and forth extremely large FetchRequests and FetchResponses, even though there is very little actual message data being added per second.  As the number of partitions grows, Kafka uses more and more network bandwidth to pass back and forth these messages.

When Kafka is tuned for lower latency, these inefficiencies become worse.  If we double the number of FetchRequests sent per second, we should expect there to be more partitions which haven't changed within the reduced polling interval.  And we are less able to amortize the high fixed cost of sending metadata for each partition in every FetchRequest and FetchResponse.  This again results in Kafka using more of the available network bandwidth.

Proposed Changes

We can solve the scalability and latency problems discussed above by creating "incremental" fetch requests that only include information about what has changed.  In order to do this, we need to introduce the concept of "fetch sessions."

Fetch Sessions

A fetch session encapsulates the state of an individual fetcher.  This allows us to avoid resending this state as part of each fetch request.

The Fetch Session includes:

1. A randomly generated 64-bit session ID which is unique on the leader
2. The client ID
   1. The numeric follower ID, if this fetch session belongs to a Kafka broker
   2. The client ID string, if this fetch session belongs to a Kafka consumer
3. The fetch epoch
4. For each partition which the fetcher is interested in:
   
   1. The topic and partition ID which uniquely identify the partition within Kafka
   2. The last fetch offset
   3. The maximum number of bytes to fetch from this partition
   4. The last dirty epoch
5. The time when the fetch session was last used

1. Fetch Session ID

The fetch session ID is a randomly generated 64-bit session ID.  It is a unique, immutable identifier for the fetch session.  Note that the fetch session ID may not be globally unique (although it's very likely to be so, since the space of 64 bit numbers is very large.)  It simply has to be unique on the leader.

Since the ID is randomly chosen, it cannot leak information to unprivileged clients.  It is also very hard for a malicious client to guess the fetch session ID.  (Of course, there are other defenses in place against malicious clients, but the randomness of the ID provides defense in depth.)

2. Client ID

The client ID describes who created this fetch session.  Sessions may be created by both brokers and Kafka clients.

It is useful to retain this ID so that we can include it in log messages, for greater debuggability.  It is also useful for cache management (more about that later.)

3. The Fetch Epoch

The fetch epoch is a monotonically incrementing 64-bit counter.  When the leader receives a fetch request with epoch N + 1, it knows that the data it sent back for the fetch request with epoch N was successfully processed by the follower.

Full fetch requests always have an epoch of 0.  Incremental fetch requests always have an epoch which is greater than 0.

4. Per-Partition Data

The FetchSession maintains information about a specific set of relevant partitions.  Note that the set of relevant partitions is established when the FetchSession is created.  It cannot be changed later.

For each partition, the last fetch offset is the latest offset which the fetcher has requested.  This is taken directly from the incremental FetchRequest RPC.  It is not updated based on the partition data that we send back to the fetcher.

We also track the maximum number of bytes to send back to the fetcher.  This value is taken directly from the last FetchRequest in the FetchSession which mentioned the partition.

A partition is considered "dirty" if it has changed and needs to be resent.  The partition becomes dirty when:

• The LogCleaner deletes messages, and this changes the log start offset of the partition on the leader.
• The leader advances the high water mark
• The leader advances the last stable offset
• The leader changes the aborted transaction list for the partition

To mark the partition as dirty, we set lastDirtyEpoch to the number of the upcoming fetch epoch.

4. The time when the fetch session was last used

This is the time in wall-clock milliseconds when the fetch session was last used.  This is used to expire fetch sessions after they have been inactive.  See the next section for details.

Fetch Session Caching

Because fetch sessions use memory on the leader, we want to limit the amount of them that we have at any given time.  Therefore, the fetch session cache has a limited number of slots.

Each broker in the cluster gets its own set of cache slots. It makes sense to reserve dedicated cache slots for brokers, because we know that brokers will be fetching lots of partitions. They should get a lot of benefit out of incremental fetch requests.  The reason for giving each broker more than one cache slot is because brokers may have several fetcher threads which operate in parallel.  In this scenario, each fetcher thread should get its own FetchSession, following a different set of partitions.

The Kafka clients all share a common pool of cache slots.  Cache slots will be allocated on a first-come, first-serve basis.  Since clients which fetch many partitions will get the most benefit out of incremental fetch requests, the server may refuse to create a FetchSession unless the client requests more than a configurable number of partitions.  This would allow clients which watch many partitions, like MirrorMaker, to get a cache slot even if there were many clients.

Fetch Sessions become eligible for eviction if they are not used within a configurable time period.  Every time the fetch session is successfully used, the timer gets reset.  In general, the time period can be fairly short-- just a minute or two.  The time period can be relatively short because most clients send fetch requests frequently.  This is certainly true for followers loading new messages from the leader.

If users have Kafka clients which consume many partitions, but take a long time between fetch requests, they may want to increase the timeout.  Alternately, they can bump up the minimum FetchSession size, so that fewer clients are competing for FetchSesssions.

Clients can voluntarily give up their FetchSessions.  They may want to do this because they want to establish a new fetch session that is monitoring a different set of partitions, or because the client is going away.

Public Interface Changes 

Error Codes

There is a new error code, InvalidFetchSession.  The server responds with this error code when it is given an fetch session ID that is no longer valid.  It may also respond with this error code if the fetch session epoch is invalid.

FetchRequest Changes

There are several changes to the FetchRequest API.

The response now contains a top-level error code.  This error code is set to indicate that the request as a whole cannot be processed.  It will be set to InvalidFetchSession if the fetch session information in the request was invalid.

New Configurations

min.fetch.session.size

The minimum size of a fetch session.  The broker will not create incremental fetch sessions for fetch requests smaller than this.

min.fetch.session.eviction.timeout.ms

Fetch sessions which are not used within this timeout are eligible for eviction.

An IncrementalFetchRequest does not enumerate the full set of partitions that the follower is interested in.  Instead, it describes only the partitions that have been updated since the last request was sent. So, for example, in our hypothetical 100,000-partition cluster, if the follower has only updated the log end offset of a single partition, it would include only that single partition in its IncrementalFetchRequest message, rather than all of them.  The corresponding IncrementalFetchResponse omits repeating metadata about partitions that haven't changed since the last IncrementalFetchRequest.

The follower registers the set of partitions which it is interested in by sending an ordinary FetchRequest.  (Because the FetchRequest currently includes a broker id, there is no ambiguity about which follower has sent the request.)  The leader then generates a random 64-bit UUID and returns it in the FetchResponse.  The follower can then make IncrementalFetchRequests with that same UUID, and the leader will understand what set of partitions the follower is interested in.  If the leader receives an IncrementalFetchRequest with a UUID that does not match that of the latest FetchResponse, it will return an error.  To resolve this error, the follower must fall back to making a full FetchRequest.

The purpose of the randomly generated UUID is to make it absolutely clear which full fetch request each incremental fetch request is based on.  For example, imagine that a follower makes two slightly different FetchRequests, and the leader ends up processing the earlier one later for some reason.  In that case, we do not want the follower and the leader to disagree about what the IncrementalFetchResponse is based on.  The UUID prevents this.  Similarly, if someone sends a full fetch request with a broker ID that is not theirs, either accidentally or maliciously, they cannot disrupt the functioning of the real broker.  The UUID, being random, cannot be forged by an attacker or counterfeited by buggy code.

This change adds a new RPC, the IncrementalFetchRequest.

IncrementalFetchRequest => uuid replica_id max_wait_time min_bytes isolation_level [topic]
  uuid => INT64
  replica_id => INT32
  max_wait_time => INT32
  min_bytes => INT32
  isolation_level => INT8
  topic => topic_name [partition]
    topic_name => STRING
    partition => partition_id fetch_offset start_offset max_bytes
      partition_id => STRING
      fetch_offset => INT64
      start_offset => INT64
      max_bytes => INT32
      
IncrementalFetchResponse => throttle_time_ms error_code uuid [topic]
  throttle_time_ms => INT32
  error_code => INT16
  uuid => INT64
  topic => topic_name [partition]
    topic_name => STRING
    partition => partition_header records
      partition_header => partition_id error_code high_watermark last_stable_offset log_start_offset [aborted_transaction]
        partition_id => INT32
        error_code => INT16
        high_watermark => INT64
        last_stable_offset => INT64
        log_start_offset => INT64
        aborted_transaction => producer_key first_offset
          producer_key => INT64
          first_offset => INT64
      records => RECORDS

This change also adds a 64-bit UUID to FetchResponse.  There are no other changes to FetchResponse.

Compatibility, Deprecation, and Migration Plan

Although this change adds a new request type, existing brokers can continue to use the old FetchRequest as before.  Therefore, there is no compatibility impact.

Rejected Alternatives

We considered several other approaches to minimizing the inefficiency of FetchRequests and FetchResponses.

• Reduce the number of temporary objects created during serialization and deserialization
• Assign each topic a unique ID, so that topic name strings do not have to be sent over the wire

Changing the way Kafka does serialization and deserialization would be a huge change, requiring rewrites of all the Request and Response classes.  We could potentially gain some efficiency by getting rid of the Builder classes for messages, but we would have to find a way to solve the problems which the Builder classes currently solve, such as the delayed binding between message content and message version and format.  Micro-optimizing the serialization and deserialization path might also involve making tradeoffs that decrease perceived code quality.  For example, we might have to sometimes give up immutability, or use primitive types instead of class types.

Assigning unique IDs to topics is also a very big project, requiring cluster-wide coordination, many RPC format changes, and potentially on-disk format changes as well.

More fundamentally, while both of these approaches improve the constant factors associated with FetchRequest and FetchResponse, they do not change the computational complexity.  In both cases, the complexity would remain O(num_extant_partitions).  In contrast, the approach above makes the messages scale as O(num_partition_changes).  Adding IncrementalFetchRequest and IncrementalFetchResponse is a smaller and more effective change.