Status

Current state: Under Discussion

Discussion thread: here

JIRA: Unable to render Jira issues macro, execution error.

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

Motivation

In KIP-500 the Active Controller, which is the quorum leader from KIP-595, will materialize the entries in the metadata log into memory. Each broker in KIP-500, which will be a voter or observer replica of the metadata log, will materialize the entries in the log into a metadata cache. We want to provide stronger guarantee on the consistency of the metadata than is possible today. If all entries in the metadata log are indefinitely retained, then the design described in KIP-595 achieves a consistent log and as a result a consistent metadata cache. If Kafka doesn’t limit the size of this replicated log it can affect Kafka availability by:

1. Running out of disk space
2. Increasing the time needed to load the replicated log into memory
3. Increasing the time it takes for new brokers to catch up to the leader's log and hence join the cluster.

Kafka uses the cleanup policies compact and delete to limit the size of the log. The delete cleanup policy deletes data for a prefix of the log based on time. The delete cleanup policy is not viable when used on metadata log because the Kafka controller and metadata cache will not be able to generate a consistent view from such a log.

The compact cleanup policy incrementally removes key/value records that are not needed to reconstruct the associated key/value state. During compaction the log cleaner keeps track of an OffsetMap which is a mapping from key to the latest offset in the section of the log being cleaned. The log cleaner assumes that any record with an offset less that the value in the OffsetMap given by the record's key is safe to remove. This solution will eventually result in one record for every key in log. This includes a tombstone record for key that have been deleted. Kafka removes this tombstone records after a configurable time. This can result in different logs in each replica which can result in different in-memory state after applying the log. We explore changes to the compact cleanup policy in the "Rejected Alternatives" section.

Type of State Machines and Operations

The type of in-memory state machines that we plan to implement for the Kafka Controller (see KIP-631) doesn't map very well to a key and offset based clean up policy. The data model for the Kafka Controller requires us to supports deltas/events. Currently the controller uses ZK for both storing the current state of the cluster and the deltas that it needs to apply. One example of this is topic deletion, when the user asks for a topic to be deleted 1) the Kafka Controller (AdminManager in the code) persists and commits the delete event, 2) applies the event to its in-memory state and sends RPC to all of the affected brokers, 3) persists and commits the changes to ZK. We don't believe that a key-value and offset based compaction model can be for events/deltas in a way that it is intuitive to program against.

Another operation that the Kafka Controller models as an event is the FenceBroker message from KIP-631. Every time a broker is fenced because of controlled shutdown or failure to send a heartbeat it affects every topic partition assigned to the fenced broker. For details on the replication impact of event operations (FenceBroker) vs key-value operation (IsrChange) see the section "Amount of Data Replicated".

Loading State and Frequency of Compaction

When starting a broker either because it is a new broker or it is restarting, loading the state represented by the __cluster_metadata and __consumer_offsets topic partition is required before the broker is available. By taking snapshots of the in-memory state as described below, Kafka can be much more efficient and aggressive on how often it generates the snapshot. By having snapshots in a separate file and not updating the replicated log the snapshot can include up to the high watermark.

For topics like __cluster_metadata and __consumer_offsets we can assume that the set of keys rarely changes but that values change frequently. With this assumption, to generate a snapshot from the in-memory state as propose in this document the broker needs to write O(the size of the data). For log compaction the broker has to write O(the size of the data) + O(the size of tombstones).

The difference is in the read pattern for the two solutions. For in-memory snapshots every log record will be read once to update the in-memory state. The Kafka Controller and the Group Coordinator will read from the head of the log. No additional reads are required to generate a snapshot.

For log compaction there are three reads:

1. One read to update the in-memory state.
2. One read to generate the map of keys to offsets.
3. One read to copy the live records from current log segments to the new log segments. These may read old offsets and may not be in the OS's file cache

If we assume an in-memory state of 100MB and a rate of change of 1MB/s, then

1. With in-memory snapshot the broker needs to read 1MB/s from the head of the log.
2. With log compaction the broker needs to
   1. read 1MB/s from the head of the log to update the in-memory state
   2. read 1MB/s to update the map of keys to offsets
   3. read 3MB/s (100MB from the already compacted log, 50MB from the new key-value records) from the older segments. The log will accumulate 50MB in 50 seconds worth of changes before compacting because the default configuration has a minimum clean ratio of 50%.

We mentioned __consumer_offsets in this section because in the future we are interested in using this mechanism for high-throughput topic partitions like the Group Coordinator and Transaction Coordinator.

Amount of Data Replicated

Because the Kafka Controller as described in KIP-631 will use events/deltas to represent some of its in-memory state changes the amount of data that needs to be written to the log and replicated to all of the brokers can be reduced. As an example if a broker goes offline and we assume that we have 100K topics each with 10 partitions with a replica factor of 3, 100 brokers. In a well balanced cluster we can expect that no broker will have 2 or more partitions for the same topic. In this case the best we can do for IsrChange message and FenceBroker message are the following message sizes:

1. IsrChange has a size of 40 bytes. 4 bytes for partition id, 16 bytes for topic id, 8 bytes for ISR ids, 4 bytes for leader id, 4 bytes for leader epoch, 2 bytes for api key and 2 bytes for api version.
2. FenceBroker has a size of 8 bytes. 4 bytes for broker id, 2 bytes for api key and 2 bytes for api version.

When a broker is shutdown and the Kafka Controller uses IsrChange to include this information in the replicated log for the __cluster_metadata topic partition then the Kafka Controller and each broker needs to persist 40 bytes per topic partitions hosted by the broker. Each broker will be a replica for 30,000 partitions, that is 1.14 MB that needs to get written in the replicated log by each broker. The Active Controller will need to replicate 1.14 MB to each broker in the cluster (99 brokers) or 113.29 MB.

When using events or deltas (FenceBroker) each broker needs to persist 8 bytes in the replicated log and 1.14 MB in the snapshot. The Active Controller will need to replica 8 bytes to each broker in the cluster (99 brokers) or .77 KB. Note that each broker still needs to persist 1.14 MB - the difference is that those 1.14 MB are not sent through the network.

Consistent Log and Tombstones

It is very important that the log remain consistent across all of the replicas after log compaction. The current implementation of log compaction in Kafka can lead to divergent logs because it is possible for slow replicas to miss the tombstone records. Tombstone records are deleted after a timeout and are not guarantee to have been replicated to all of the replicas before they are deleted. It should be possible to extend log compaction and the fetch protocol as explained in the Rejected Alternatives "Just fix the tombstone GC issue with log compaction". We believed that this conservative approach is equally, if not more complicated than snapshots as explain in this proposal and more importantly it doesn't address the other items in the motivation section above.

Overview

This KIP assumes that KIP-595 has been approved and implemented. With this improvement replicas will continue to replicate their log among each others. Replicas will snapshot their log independent of each other. Snapshots will be consistent because the logs are consistent across replicas. Follower and observer replicas fetch the snapshots from the leader when they attempt to fetch an offset from the leader and the leader doesn’t have that offset in the log.

Generating and loading the snapshot will be delegated to the Kafka Controller. The Kafka Controller will notify the Kafka Raft layer when it has generated a snapshot and up to which offset is included in the snapshot. The Kafka Raft layer will notify the Kafka Controller when a new snapshot has been fetched from the Active Controller. See section "Interaction Between Kafka Raft and the Kafka Controller" and KIP 595 for more details.

This KIP focuses on the changes required for use by the Kafka Controller's __cluster_metadata topic partition but we should be able to extend it to use it in other internal topic partitions like __consumer_offsets and __transaction.

Proposed Changes

The __cluster_metadata topic will have snapshot as the cleanup.policy. The Kafka Controller will have an in-memory representation of the replicated log up to at most the high-watermark. When performing a snapshot the Kafka Controller will serialize this in-memory state to disk. This snapshot file on disk is described by the largest offset and epoch from the replicated log that has been included.

The Kafka Controller will notify the Kafka Raft client when it has finished generating a new snapshots. It is safe to truncate a prefix of the log up to the latest snapshot. The __cluster_metadata topic partition will have the latest snapshot and zero or more older snapshots. Having multiple snapshots is useful for minimizing re-fetching of the snapshot when a new snapshot is generated. These additional snapshots would have to be deleted, this is described in "When to Delete Snapshots". 

Design

Visually a Kafka Raft Log would look as follow:

Kafka Replicated Log:

      LBO  --               high-watermark --      LEO --
            V                               V           V
        ---------------------------------------------------
offset: | x + 1 | ... | y | y + 1 |  ...  |   |  ...  |   |
epoch:  | b     | ... | c | d     |  ...  |   |  ...  |   |
        ---------------------------------------------------

Kafka Snapshot Files:

<topic_name>-<partition_index>/x-a.checkpoint
<topic_name>-<partition_index>/y-c.checkpoint

Note: The extension checkpoint will be used since Kafka already has file with the snapshot extension.

Legends

1. LEO - log end offset - the largest offset and epoch that has been written to disk.

2. high-watermark - the largest offset and epoch that has been replicated to N/2 + 1 replicas.

3. LBO - log begin offset - the smallest offset in the replicated log. This property used to be called log start offset. It is renamed such that the acronym doesn’t conflict with last stable offset (LSO).

In the example above, offset=x, epoch=a does not appear in the replicated log because it is before the LBO (x + 1, b). If the Kafka topic partition leader receives a fetch request with an offset and epoch greater than or equal to the LBO (x + 1, b) then the leader handles the request as describe in KIP-595. Otherwise, the leader will respond with the offset and epoch of the latest snapshot (y, c) and with the next fetch offset and epoch (y + 1, d).

The followers and observers can concurrently fetch both the snapshot and the replicated log. During leader election, followers with incomplete or missing snapshot will send a vote request and response as if they had an empty log.

Snapshot Format

The Kafka Controller is responsible for the format of the snapshot. The snapshot will be stored in the topic partition directory and will have a name of <SnapshotOffset>-<SnapshotOffsetEpoch>.checkpoint. For example for the topic __cluster_metadata, partition 0, snapshot offset 5120793 and snapshot epoch 2, the full filename will be __cluster_metadata-0/00000000000005120793-2.checkpoint.

For details on the content of this snapshot file and how it manages compatibility with future changes see KIP-631 and the section on Record Format Versions.

When to Snapshot

If the Kafka Controller generates a snapshot too frequently then can negatively affect the performance of the disk. If it doesn't generate a snapshot often enough then it can increase the amount of time it takes to load its state into memory and it can increase the amount space taken up by the replicated log. The Kafka Controller will have a new configuration option controller.snapshot.minimum.records.  If the number of records between the latest snapshot and the high-watermark is greater than controller.snapshot.minimum.records, then the Kafka Controller will perform a new snapshot.

When to Increase the Log Begin Offset

To minimize the need to send snapshots to slow replicas, the leader can delay increasing the log begin offset up to the minimum offset that it knows it has been replicate to all of the live replicas. The leader's LBO can be increased to an offset X if the following is true:

1. There is a snapshot for the topic partition which includes the offset for X - 1 and.

2. One of the following is true:

   1. All of the live replicas (followers and observers) have replicated LBO or

   2. LBO is max.replication.lag.ms old.

Followers and observers will increase their log begin offset to the value sent on the fetch response as long as the local state machine has generated a snapshot that includes to the log begin offset minus one.

Kafka allows the clients to delete records that are less than a given offset by using the DeleteRecords RPC . Those requests will be validated using the same logic enumerated above.

When to Delete Snapshots

The broker will delete any snapshot with a latest offset and epoch (SnapshotOffsetAndEpoch) less than the LBO - 1. The means that the question of when to delete a snapshot is the same as when to increase the LBO which is covered the the section “When To Increase the Log Begin Offset”.

Interaction Between Kafka Raft and the Kafka Controller

This section describe the interaction between the state machine and Kafka Raft.

From Kafka Raft to the Kafka Controller

1. Notify when a snapshot is available because of a FetchSnapshot.

From the Kafka Controller to Kafka Raft

1. Notify when snapshot finished. This includes the largest included offset and epoch.

Loading Snapshots

There are two cases when the Kafka Controller needs to load the snapshot:

1. When it is booting.

2. When the follower and observer replicas finishes fetching a new snapshot from the leader.

After loading the snapshot, the Kafka Controller will apply operations in the replicated log with an offset greater than the snapshot’s offset and less than the high-watermark.

Changes to Leader Election

The followers of the __cluster_metadata topic partition will concurrently fetch the snapshot and replicated log. This means that candidates with incomplete snapshots will send a vote request with a LastOffsetEpoch of -1 and a LastOffset of -1 no matter the LEO of the replicated log.

Validation of Snapshot and Log

There is an invariant that every log must have at least one snapshot where the offset of the snapshot is between LogBeginOffset - 1 and High-Watermark. If this is not the case then the replica should assume that the log is empty. Once the replica discovers the leader for the partition it can bootstrap itself by using the Fetch and FetchSnapshot RPCs as described in this document.

Public Interfaces

Topic configuration

1. The __cluster_metadata topic will have an cleanup.policy value of snapshot. This configuration can only be read. Updates to this configuration will not be allowed.

2. controller.snapshot.minimum.records is the minimum number of records committed after the latest snapshot before the Kafka Controller will perform another snapshot. See section "When to Snapshot"

3. max.replication.lag.ms - The maximum amount of time that leader will wait for an offset to get replicated to all of the live replicas before advancing the LBO. See section “When to Increase the Log Begin Offset”.

Network Protocol

Fetch

Request Schema

This KIP doesn’t require any changes to the fetch request outside of what is proposed in KIP-595.

Response Schema

The fetch response proposed in KIP-595 will be extended such that:

1. It includes new optional fields for snapshot offset and snapshot epoch.

2. The LogStartOffset field is renamed to LogBeginOffset.

{
  "apiKey": 1,
  "type": "response",
  "name": "FetchResponse",
  "validVersions": "0-12",
  "flexibleVersions": "12+",
  "fields": [
    { "name": "ThrottleTimeMs", "type": "int32", "versions": "1+", "ignorable": true,
      "about": "The duration in milliseconds for which the request was throttled due to a quota violation, or zero if the request did not violate any quota." },
    { "name": "ErrorCode", "type": "int16", "versions": "7+", "ignorable": false,
      "about": "The top level response error code." },
    { "name": "SessionId", "type": "int32", "versions": "7+", "default": "0", "ignorable": false,
      "about": "The fetch session ID, or 0 if this is not part of a fetch session." },
    { "name": "Topics", "type": "[]FetchableTopicResponse", "versions": "0+",
      "about": "The response topics.", "fields": [
      { "name": "Name", "type": "string", "versions": "0+", "entityType": "topicName",
        "about": "The topic name." },
      { "name": "Partitions", "type": "[]FetchablePartitionResponse", "versions": "0+",
        "about": "The topic partitions.", "fields": [
        { "name": "PartitionIndex", "type": "int32", "versions": "0+",
          "about": "The partiiton index." },
        { "name": "ErrorCode", "type": "int16", "versions": "0+",
          "about": "The error code, or 0 if there was no fetch error." },
        { "name": "HighWatermark", "type": "int64", "versions": "0+",
          "about": "The current high water mark." },
        { "name": "LastStableOffset", "type": "int64", "versions": "4+", "default": "-1", "ignorable": true,
          "about": "The last stable offset (or LSO) of the partition. This is the last offset such that the state of all transactional records prior to this offset have been decided (ABORTED or COMMITTED)" },
        // Field renamed from LogStartOffset to LogBeginOffset 
        { "name": "LogBeginOffset", "type": "int64", "versions": "5+", "default": "-1", "ignorable": true,
          "about": "The current log begin offset." },
        // ---------- Start new fields by KIP-595 ----------
        { "name": "NextOffsetAndEpoch", "type": "OffsetAndEpoch",
          "versions": "12+", "taggedVersions": "12+", "tag": 0, "fields": [
          { "name": "NextFetchOffset", "type": "int64", "versions": "0+",
            "about": "If set, this is the offset that the follower should truncate to"},
          { "name": "NextFetchOffsetEpoch", "type": "int32", "versions": "0+",
            "about": "The epoch of the next offset in case the follower needs to truncate"}
        ]},
        { "name": "CurrentLeader", "type": "LeaderIdAndEpoch",
          "versions": "12+", "taggedVersions": "12+", "tag": 1, fields": [
          { "name": "LeaderId", "type": "int32", "versions": "0+",
            "about": "The ID of the current leader or -1 if the leader is unknown."},
          { "name": "LeaderEpoch", "type": "int32", "versions": "0+",
            "about": "The latest known leader epoch"}
        ]},
        // ---------- End new fields by KIP-595 ----------
        // ---------- Start new fields ----------
        { "name": "SnapshotOffsetAndEpoch", "type": "OffsetAndEpoch",
          "versions": "12+", "taggedVersions": "12+", "tag": 0, "fields": [
          { "name": "FetchOffset", "type": "int64", "versions": "0+",
            "about": "If set, this is the offset that the follower should use for the FetchSnapshot request"},
          { "name": "FetchOffsetEpoch", "type": "int32", "versions": "0+",
            "about": "If set, this is the epoch that the follower should use for the FetchSnapshot request"}
        ]},
        // ---------- End new fields ----------
        { "name": "Aborted", "type": "[]AbortedTransaction", "versions": "4+", "nullableVersions": "4+", "ignorable": false,
          "about": "The aborted transactions.",  "fields": [
          { "name": "ProducerId", "type": "int64", "versions": "4+", "entityType": "producerId",
            "about": "The producer id associated with the aborted transaction." },
          { "name": "FirstOffset", "type": "int64", "versions": "4+",
            "about": "The first offset in the aborted transaction." }
        ]},
        { "name": "PreferredReadReplica", "type": "int32", "versions": "11+", "ignorable": true,
          "about": "The preferred read replica for the consumer to use on its next fetch request"},
        { "name": "Records", "type": "bytes", "versions": "0+", "nullableVersions": "0+",
          "about": "The record data." }
      ]}
    ]}
  ]
}

Handling Fetch Request

The leader’s handling of fetch request will be extended such that if FetchOffset is less than LogBeginOffset then the leader will respond with a SnapshotOffsetAndEpoch of the latest snapshot and a NextOffsetAndEpoch of the offset right after SnapshotOffsetAndEpoch.

Handling Fetch Response

The replica's handling of fetch response will be extended such that if SnapshotOffsetAndEpoch is set then the follower will delete its entire log and snapshots, and start fetching from the leader both the log and the snapshot. Where to fetch in the Log is describe by NextOffsetAndEpoch while which snapshot to fetch is described by SnapshotOffsetAndEpoch.

Fetching Snapshots

Request Schema

{
  "apiKey": #,
  "type": "request",
  "name": "FetchSnapshotRequest",
  "validVersions": "0",
  "flexibleVersions": "0+",
  "fields": [
    { "name": "ReplicaId", "type": "int32", "versions": "0+",
      "about": "The broker ID of the follower, of -1 if this request is from a consumer." },
    { "name": "Topics", "type": "[]TopicSnapshots", "versions": "0+",
      "about": "The topics to fetch.", "fields": [
      { "name": "Name", "type": "string", "versions": "0+", "entityType": "topicName",
        "about": "The name of the topic to fetch." },
      { "name": "Partitions", "type": "[]PartitionSnapshots", "versions": "0+",
        "about": "The partitions to fetch.", "fields": [
        { "name": "PartitionIndex", "type": "int32", "versions": "0+",
          "about": "The partition index." },
        { "name": "SnapshotOffsetAndEpoch", "type": "OffsetAndEpoch",
          "versions": "12+", "taggedVersions": "12+", "tag": 0, "fields": [
          { "name": "Offset", "type": "int64", "versions": "0+",
            "about": "If set, this is the offset that the follower should use for the FetchSnapshot request"},
          { "name": "OffsetEpoch", "type": "int32", "versions": "0+",
            "about": "If set, this is the epoch that the follower should use for the FetchSnapshot request"}
        ]},
        { "name": "Position", "type": "int64", "versions": "0+",
          "about": "The byte position within the snapshot." },
        { "name": "MaxBytes", "type": "int32", "versions": "0+",
          "about": "The maximum bytes to fetch from this snapshot." }
      ]}
    ]}
  ]
}

Response Schema

{
  "apiKey": #,
  "type": "request",
  "name": "FetchSnapshotResponse",
  "validVersions": "0",
  "flexibleVersions": "0+",
  "fields": [
    { "name": "ThrottleTimeMs", "type": "int32", "versions": "0+", "ignorable": true,
      "about": "The duration in milliseconds for which the request was throttled due to a quota violation, or zero if the request did not violate any quota." },
    { "name": "ErrorCode", "type": "int16", "versions": "0+", "ignorable": false,
      "about": "The top level response error code." },
    { "name": "Topics", "type": "[]TopicSnapshots", "versions": "0+",
      "about": "The topics to fetch.", "fields": [
      { "name": "Name", "type": "string", "versions": "0+", "entityType": "topicName",
        "about": "The name of the topic to fetch." },
      { "name": "Partitions", "type": "[]PartitionSnapshots", "versions": "0+",
        "about": "The partitions to fetch.", "fields": [
        { "name": "PartitionIndex", "type": "int32", "versions": "0+",
          "about": "The partition index." },
        { "name": "ErrorCode", "type": "int16", "versions": "0+",
          "about": "The error code, or 0 if there was no fetch error." },
	    { "name": "Continue", "type": "bool": "versions": "0+",
	      "about": "True if there is data remaining to fetch." },
        { "name": "Position", "type": "int64", "versions": "0+",
          "about": "The byte position within the snapshot." },
	    { "name": "Bytes", "type": "bytes": "versions": "0+",
	      "about": "Snapshot data." },
      ]}
    ]}
  ]
}

Request Handling

When the leader receives a FetchSnapshot request it would do the following:

1. Find the snapshot for SnapshotOffsetAndEpoch for the topic Name and partition PartitionIndex.

2. Send the bytes in the snapshot from Position up to at most MaxBytes.

3. Set Continue to true if there are renaming bytes in the snapshots.

Errors:

1. SNAPSHOT_NOT_FOUND - when the fetch snapshot request specifies a SnapshotOffsetAndEpoch that doesn’t exists on the leader.

2. OFFSET_OUT_OF_RANGE - when the fetch snapshot request’s offset is greater than the size of the snapshot.

3. NOT_LEADER_FOR_PARTITION - when the fetch snapshot request is sent to a replica that is not the leader.

Response Handling

1. Copy the bytes in Bytes to the local snapshot file starting at file position Position.

2. If Continue is true, the follower is expected to send another FetchSnapshotRequest with a Position set to the response's Position plus the response's len(Bytes).

3. If Continue is false then fetching the snapshot finished. Notify the Kafka Controller that a new snapshot is available. The Kafka Controller will load this new snapshot and read records from the log after the snapshot offset.

Errors:

1. SNAPSHOT_NOT_FOUND, OFFSET_OUT_OF_RANGE, NOT_LEADER_FOR_PARTITION - The follower can rediscover the new snapshot by sending a fetch request with an offset and epoch of zero.

Metrics

Compatibility, Deprecation, and Migration Plan

This KIP is only implemented for the internal topic __cluster_metadata. Internal and external clients for all other topics can ignore the SnapshotOffsetAndEpoch as that field will not be set for those topic partitions.

Rejected Alternatives

Append the snapshot to the Kafka log: Instead of having a separate file for snapshots and separate RPC for downloading the snapshot, the leader could append the snapshot to the log. This design has a few drawbacks. Because the snapshot gets appended to the head of the log that means that it will be replicated to every replica in the partition. This will cause Kafka to use more network bandwidth than necessary.

Use Kafka Log infrastructure to store snapshots: Kafka could use the Log functionality to persists the snapshot to disk and to facilitate the replication of the snapshot to the replicas. This solution ties the snapshot format to Kafka Log’s record format. To lower the overhead of writing and loading the snapshot by the Kafka Controller, it is important to use a format that is compatible with the in-memory representation of the Kafka Controller.

Change leader when generating snapshots: Instead of supporting concurrent snapshot generation and writing to the in-memory state, we could require that only non-leaders generate snapshots. If the leader wants to generate a snapshot then it would relinquish leadership. We decided to not implement this because:

1. All replicas/brokers in the cluster will have a metadata cache which will apply the replicated log and allow read-only access by other components in the Kafka Broker.
2. Even though leader election within the replicas in the replicated log may be efficient, communicating this new leader to all of the brokers and all of the external clients (admin clients) may have a large latency.
3. Kafka wants to support quorums where there is only one voter replica for the Kafka Controller.

Just fix the tombstone GC issue with log compaction: we have also talked about just fixing the tombstone garbage collecting issue as listed in motivation instead of introducing the snapshot mechanism. The key idea is that fetchers either from follower or from consumer needs to be notified if they "may" be missing some tombstones that have deprecated some old records. Since the offset span between the deprecated record and the tombstone for the same key can be arbitrarily long — for example, we have a record with key A at offset 0, and then a tombstone record for key A appended a week later, at offset 1 billion — we'd have to take a conservative approach here. More specifically:

• The leader would maintain an additional offset marker up to which we have deleted tombstones, i.e. all tombstones larger than this offset are still in the log. Let's call it a tombstone horizon marker. This horizon marker would be returned in each fetch response.
• Follower replicas would not do delete its tombstones beyond the received leader's horizon marker, and would maintain it's own marker as well which should always be smaller than leader's horizon marker offset.
• All fetchers (follower and consumer) would keep track of the horizon marker from the replica it is fetching from. The expectation is that leaders should only delete tombstones in a very lazy manner, controlled by the existing config, so that the fetching position should be larger than the horizon marker most of the time. However, if the fetcher does found its fetching position below the horizon marker, then it needs to handle it accordingly:
  • For consumer fetchers, we need to throw an extended error of InvalidOffsetException to notify the users that we may missed some tombstones, and let the users to reset is consumed record states — for example, if the caller of consumer is a streams client, and it is building a local state from the fetched records, that state need to be wiped.
  • For replica fetchers, we need to wipe the local log all together and then restart fetching from the starting offset.
  • For either types of fetchers, we will set an internal flag, let's call it unsafe-bootstrapping mode to true, while remembering the horizon marker offset at the moment from the replica it is fetching from.
    • When this flag is set to true, we would not fall back to a loop of this handling logic even when the fetching position is smaller than the horizon marker, since otherwise we would fall into a loop of rewinding.
    • However, during this period of time, if the horizon marker offsets get updated from the fetch response, meaning that some tombstones are deleted again, then we'd need to retry this unsafe-bootstrap from step one.
    • After the fetching position as advanced beyond the remembered horizon marker offset, we would reset this flag to false to go back to the normal mode.

We feel this conservative approach is equally, if not more complicated than adding a snapshot mechanism since it requires rather tricky handling logic on the fetcher side; plus it only resolves one of the motivations — the tombstone garbage collection issue.

Another thought against this idea is that for now most of the internal topics (controller, transaction coordinator, group coordinator) that would need the snapshot mechanism describe in this KIP can hold their materialized cache in memory completely, and hence generating snapshots should be comparably more efficient to execute.