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 DiscussionAccepted

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

...

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.

...

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

...

The Fetch Session includes:

1. A randomly generated 6432-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
   The 32-bit fetch epoch
3. For Cached data about each partition which the fetcher is interested in:
   .
4. The privileged bit
5. The topic and partition ID which uniquely identify the partition within Kafka
6. The last fetch offset
7. The maximum number of bytes to fetch from this partition
The last dirty epoch
8. The time when the fetch session was last used

...

Fetch Session ID

The fetch session ID is a randomly generated 6432-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

Fetch Epoch

The fetch epoch is a monotonically incrementing 32-bit counter. After processing request N, the broker expects to receive request N+1.

The sequence number is always greater than 0.  After reaching MAX_INT, it wraps around to 1.

Cached data about each partition

If the fetch session supports incremental fetches, the FetchSession will maintain information about each partition in the incremental fetch.

For each partition, we maintain:

• The topic name
• The partition index
• The maximum number of bytes to fetch from this partition
• The fetch offset
• The high water mark
• The fetcher log start offset
• The leader log start offset

Topic name and partition index come from the TopicPartition.

maxBytes, fetchOffset, and fetcherLogStartOffset come from the latest FetchRequest in which the partition appeared.

highWatermark and localLogStartOffset come from the leader.

The leader uses this cached information to decide which partitions to include in the FetchResponse.  Whenever any of these elements change, or if there is new data available for a partition, the partition will be included.

Privileged bit

The privileged bit is set if the fetch session was created by a follower.  It is cleared if the fetch session was created by a regular consumer.

This is retained in order to prioritize followers over consumers, when resources are low.  See the section on fetch session caching for details.

The time when the fetch session was last used

This is the time in wallThis 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 on fetch session caching 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 each broker will create only 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.

...

incremental fetch sessions.

There is one new public configurations for fetch session caching:

• max.incremental.fetch.session.cache.slots, which set the number of incremental fetch session cache slots on the broker.  Default value: 1,000

There is one new constant for fetch session caching:

• min.incremental.fetch.session.eviction.ms, which sets the minimum amount of time we will wait before evicting an incremental fetch session from the cache.  Value: 120,000

When the server gets a new request to create an incremental fetch session, it will compare the proposed new session with its existing sessions.  The new session will evict an existing session if and only if:

1. The new session belongs to a follower, and the existing session belongs to a regular consumer, OR
2. The existing session has been inactive for more than min.incremental.fetch.session.eviction.ms, OR
   
3. The existing session has existed for more than min.incremental.fetch.session.eviction.ms, AND the new session has more partitions

This accomplishes a few different goals:

• Followers get priority over consumers
• Inactive session get replaced over time
• Bigger requests, which benefit more from being incremental, are prioritized
• Cache thrashing is limited, avoiding expensive session re-establishment when there are more fetchers than cache slots.

Public Interface Changes 

New Error Codes

FetchSessionIdNotFound: The server responds with this error code when the client request refers to a fetch session that the server does not know about.  This may happen if there was a client error, or if the fetch session was evicted by the server.

InvalidFetchSessionEpochException.  The server responds with this error code when the fetch session epoch of a request is different than what it expected.

FetchRequest Changes

There are several changes to the FetchRequest API.

Fetch Session ID

A 32-bit number which identifies the current fetch session.  If this is set to 0, there is no current fetch session.

Fetch Session Epoch

A 32-bit number which identifies the current fetch session epoch.  Valid session epochs are always positive-- they are never 0 or negative numbers.

The fetch session epoch is incremented by one for each fetch request that we send.  Once it reaches MAX_INT, the next epoch is 1.

The fetch epoch keeps the state on the leader and the follower synchronized.  It ensures that if a message is duplicated or lost, the server will always notice.  It is also used to associate requests and responses in the logs. Other numbers, such as IP addresses and ports, or NetworkClient sequence numbers, can also be helpful for this purpose-- but they are less likely to be unique.

FetchRequest Metadata meaning

Incremental Fetch Requests

Incremental fetch requests have a positive fetch session ID.

A partition is only included in an incremental FetchRequest if:

• The client wants to notify the broker about a change to the partition's maxBytes, fetchOffset, or logStart
• The partition was not included in the incremental fetch session before, but the client wants to add it.
• The partition is in the incremental fetch session, but the client wants to remove it.

If the client doesn't want to change anything, the client does not need to include any partitions in the request at all.

If the client wants to remove a partition, the client will add the partition to the removed_partition list in the relevant removed_topics entry.

Schema

FetchRequest => max_wait_time replica_id min_bytes isolation_level fetch_session_id fetch_session_epoch [topic] [removed_topic]

  max_wait_time => INT32

  replica_id => INT32

  min_bytes => INT32

  isolation_level => INT8

  fetch_session_id => INT32

  fetch_session_epoch => INT32

  topic => topic_name [partition]

  topic_name => STRING

  partition => partition_id fetch_offset start_offset max_bytes

  partition_id => INT32

  fetch_offset => INT64

  start_offset => INT64

  max_bytes => INT32

  removed_topic => removed_topic_name [removed_partition_id]

  removed_topic_name => STRING

  removed_partition_id => INT32

FetchResponse Changes

Top-level error code

Per-partition error codes are no longer sufficient to handle all response errors.  For example, when an incremental fetch session encounters an FetchSessionIdNotFoundException, we do not know which partitions the client expected to fetch.  Therefore, the FetchResponse now contains a top-level error code.  This error code is set to indicate that the request as a whole cannot be processed.

When the top-level error code is set, the caller should assume that all the partitions in the fetch received the given error.

Fetch Session ID

The FetchResponse now contains a 32-bit fetch session ID.

FetchResponse Metadata meaning

Incremental Fetch Responses

A partition is only included in an incremental FetchResponse if:

• The broker wants to notify the client about a change to the partition's highWatermark or broker logStartOffset
• There is new data available for a partition

If the broker has no new information to report, it does not need to include any partitions in the response at all.

The format of the partition data within FetchResponse is unchanged.

Handling Partition Size Limits in Incremental Fetch Responses

Sometimes, the per-fetch-request limit is too small to allow us to return information about every partition.  In those cases, we will limit the number of partitions that we return information about, to avoid exceeding the per-request maximum.  (As specified in KIP-74, the response will always return at least one message, though.)

If we always returned partition information in the same order, we might end up "starving" the partitions which came near the end of the order.  With full fetch requests, the client can rotate the order in which it requests partitions in order to avoid this problem.  However, incremental fetch requests need not explicitly specify all the partitions.  Indeed, an incremental fetch request may contain no partitions at all.

In order to solve the starvation problem, the server must rotate the order in which it returns partition information.  The server does this by maintaining a linked list of all partitions in the fetch session.  When data is returned for a partition, that partition is moved to the end of the list.  This ensures that we eventually return data about all partitions for which data is available.

Schema

FetchResponse => throttle_time_ms error_code error_string fetch_session_id [topic]

  throttle_time_ms => INT32

  error_code => INT16

  fetch_session_id => INT32

  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

New Metrics

The following new metrics will be added to track cache consumption:

NumIncrementalFetchSessions: Tracks the number of incremental fetch sessions which exist.

NumIncrementalFetchPartitionsCached: Tracks the total number of partitions cached by incremental fetch sessions.

IncrementalFetchSessionEvictionsPerSec: Tracks the number of incremental fetch sessions that were evicted from the cache per second. This metric is not increased when a client closes its own incremental fetch session

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 typethe concept of incremental fetch requests, existing brokers can continue to use the old FetchRequest as before.  Therefore, there is no compatibility impact.

...