Status

Current state: Under DiscussionAdopted

Discussion thread: Thread

JIRA:

...

As described in KIP-844, under EOS, crash failures cause all Task state to be wiped out on restart. This is because, currently, data is written to the StateStore before the commit to its changelog has completed, so it's possible that records are written to disk that were not committed to the store changelog.

This ensures consistency of local stores with their changelog topics, but can cause long delays in processing while it rebuilds the local state from the changelog. These delays are proprotional to the number of records in the changelog topic, which for highly active tables, or those with a very high cardinality, can be very large. Real-world use-cases have been observed where these delays can span multiple days, where both processing, and interactive queries, are paused.

In KIP-844, it was proposed to create an alternative type of StateStore, which would enable users to opt-in to "transactional" behaviour, that ensured data was only persisted once the changelog commit has succeeded. However, the design and approach outlined in KIP-844 unfortunately did not perform well when In KIP-844, it was proposed to create an alternative type of StateStore, which would enable users to opt-in to "transactional" behaviour, that ensured data was only persisted once the changelog commit has succeeded. However, the design and approach outlined in KIP-844 unfortunately did not perform well when tested (with a write throughput that was approximately only 4% of the regular RocksDB StateStore!).

This KIP explores an alternative design that should have little/no performance impact, potentially performing better than the status quo, and can thus be enabled for all stores.

Public Interfaces

Changed:

• org.apache.kafka.streams.processor.StateStore

Changes:

This should bound state restore under EOS to less than 1 second, irrespective of the size of the changelogs.

Public Interfaces

New configuration

Changed Interfaces

• org.apache.kafka.streams.processor.StateStore
• org.apache.kafka.streams.processor.StateStoreContext

Changes:

   
    /**

    /**
     * Flush any cached data
     *
     * @deprecated since KIP-NEXT. Use {@link #commit(Map)} instead.
     */
    @Deprecated
    default void flush() {}

    /**
     * Commit the current transaction to this StateStore with the given changelog offset.
     * <p>
     * This is a convenience method for {@link #commit(Map) commit(Collections.singletonMap(null, changelogOffset))}.
     * <p>
     * This method is used to commit records to a regular, logged StateStore.
     *
     * @see #commit(Map)Flush any cached data
     * @param
 changelogOffset The offset of the* changelog@deprecated topicUse this commit corresponds to. The offset can be{@link org.apache.kafka.streams.processor.api.ProcessingContext#commit() ProcessorContext#commit()}
     *             instead.
     */
    @Deprecated
    {@code null} if this StateStore does not have a changelog.
     */
    @Evolving
    default void commit(final Long changelogOffsetdefault void flush() {
        commit(Collections.singletonMap(null, changelogOffset));// no-op
    }

    /**
     * Commit theall currentwritten transactionrecords to this StateStore with the given offsets.
     * <p>
     * AllThis recordsmethod thatMUST wereNOT writtenbe tocalled thisby storeusers sincefrom the last {@code commit} will be written to the store
     * atomically, i.e. either all will be written, or none of them will{@link org.apache.kafka.streams.processor.api.Processor processors},
     * as doing so may violate the consistency guarantees provided by this store, and expected by Kafka Streams.
     * <p>
Instead, users should call  * After this method returns, calls to {@link #getCommittedOffset(TopicPartition)} will return the associated value{@link org.apache.kafka.streams.processor.api.ProcessingContext#commit() 
     * ProcessorContext#commit()} to request a Task commit.
     * of {@code offsets}.<p>
     * <p>
When called, every write written *since {@codethe offsets}last willcall containto one of either:
     * <ul>{@link #commit(Map)}, or since this store was {@link
     * #init(StateStoreContext, StateStore) opened} will <li>A {@code Map} of offsets for each input partition of this Global StateStore.</li>be made available to readers using the {@link
     * org.apache.kafka.common.IsolationLevel#READ_COMMITTED READ_COMMITTED} {@link
     * org.apache.kafka.common.IsolationLevel IsolationLevel}.
     <li>A* {@code Map} containing one mapping from {@code null} to the offset of the changelog partition for this
     *<p>
     * If {@link #persistent()} returns {@code true}, after this method returns, all records written since the last call
     logged StateStore.</li>
     *     <li>A {@code null}, if this StateStore is not {@link StoreBuilder#withLoggingDisabled() logged}.</li>
     * </ul>* to {@link #commit(Map)} are guaranteed to be persisted to disk, and available to read, even if this {@link
     * StateStore} is {@link #close() closed} and subsequently {@link #init(StateStoreContext, StateStore) re-opened}.
     * <p>
     * If this store is {@link #persistent#managesOffsets() persistent}, it<em>also</em> isreturns required that on-restart{@code true}, the offsets returned by {@linkgiven {@code changelogOffsets} will be
     * #getCommittedOffset(TopicPartition)} correspond with the records persisted in the StateStoreguaranteed to be persisted to disk along with the written records.
     * <p>
     * @param{@code offsetschangelogOffsets} The offset(s) for the input topics this commit corresponds to. May be {@code null} if this
     *           will usually contain a single partition, in the case of a regular StateStore. However,
     * they may contain multiple partitions in the case of a Global StateStore with multiple partitions. All provided
     StateStore* haspartitions no<em>MUST</em> inputbe topicspersisted orto changelogdisk.
     */ <p>
    @Evolving
 * Implementations <em>SHOULD</em> defaultensure void commit(final Map<TopicPartition, Long> offsets) {
    that {@code changelogOffsets} are committed to disk atomically with the
    flush();
 * records they }represent.

     /** 
     * Returns@param thechangelogOffsets latest committedThe changelog offset(s) corresponding thatto the materializedmost staterecently in this store representswritten records.
     * <p>/
    default *void The state in the store is guaranteed to reflect exactly the state in the changelog up to this offset.commit(final Map<TopicPartition, Long> changelogOffsets) {
        flush();
    }

     /** <p>
     * ThisReturns the maymost berecently {@code null}, if@link #commit(Map) committed} offset for the storegiven does not contain any metadata on its changelog offset. In this case, the
     * StateStore itself makes no guarantees about its contents.{@link TopicPartition}.
     * <p>
     * If {@link #managesOffsets()} and {@link #persistent()} both return {@code true}, this method will return the
     * <p>
     * Equivalent to calling {@code getCommittedOffset(null)offset that corresponds to the changelog record most recently written to this store, for the given {@code
     * partition}.
     * <p>
     * This method shouldprovides <em>not</em>readers beusing calledthe for global StateStores, as they are not backed by a changelog.
     *{@link org.apache.kafka.common.IsolationLevel#READ_COMMITTED} {@link
     * org.apache.kafka.common.IsolationLevel} a means to determine the point in the changelog that this StateStore
     * @returncurrently Therepresents.
 latest committed changelog offset that* the
 state in this store represents, or {@code null}, if no
     ** @param partition The partition to get the committed offset for.
     * @return The last  changelog{@link #commit(Map) committed} offset metadata is available and no guarantees can be made.for the {@code partition}; or {@code null} if no offset
     */
     @Evolving
    defaulthas Longbeen getCommittedOffset() {
        return getCommittedOffset(null);
    }

committed for the partition, or if either {@link #persistent()} or {@link #managesOffsets()}
     /**
     * Returns the latest committedreturn changelog offset that the materialized state in this store represents.
     * <p>{@code false}.
     */
    default Long committedOffset(final TopicPartition partition) {
     *   Thereturn statenull;
 in the store is}

 guaranteed to reflect exactly/**
 the state in the changelog* upDetermines toif this offset StateStore manages its own offsets.
     * <p>
     * ThisIf this maymethod bereturns {@code nulltrue}, ifthen theoffsets storeprovided doesto not contain any metadata on its changelog offset. In this case, the
     * StateStore itself makes no guarantees about its contents{@link #commit(Map)} will be retrievable using
     * {@link #committedOffset(TopicPartition)}, even if the store is {@link #close() closed} and later re-opened.
     * <p>
     * If {@codethis topicPartition}method isreturns {@code nullfalse}, offsets the returned offsetprovided to {@link #commit(Map)} will be theignored, offsetand for{@link
 the changelog partition of
 * #committedOffset(TopicPartition)} will be *expected thisto StateStore,always ifreturn one{@code existsnull}.
     * <p>
     * @returnThis Themethod latestis committedprovided to offsetenable thatcustom theStateStores stateto opt-in thisto storemanaging represents,their orown {@code null}, if nooffsets. This is highly
     * recommended, if possible, to ensure that custom StateStores changelogprovide offset metadata is available and no guarantees can be made.the consistency guarantees that Kafka Streams
     */
 expects when  @Evolving
    default Long getCommittedOffset(final TopicPartition topicPartition) {operating under the {@code exactly-once} {@code processing.mode}.
     * 
     * @return Whether this StateStore manages its own offsets.
     */
   return null; default boolean managesOffsets() {
    }

Proposed Changes

There are two parts to this KIP:

1. Buffering writes to RocksDB using WriteBatchWithIndex.
2. Moving responsibility for store checkpointing inside the StateStore  itself.

Buffering writes with WriteBatchWithIndex

RocksDB provides WriteBatchWithIndex as a means to accomplishing atomic writes when not using the RocksDB WAL. The performance overhead of doing this should be negligible. The only performance consideration is that the buffer must reside completely in-memory until it is committed. RocksDB recommend buffers no more than 3-4 MiB for optimal performance. With a commit.interval.ms  of 100 milliseconds, which is the default when under EOS, and an average record size of 1KiB, a 4MiB buffer should allow for a throughput of ~40,960 records/second. This is a worst-case estimate, as most use-cases will have a considerably smaller record size, providing for significantly increased throughput. For use-cases with larger record sizes, higher throughput could be sustained, at the cost of a higher memory usage.

It is therefore not generally expected that this will cause any out-of-memory errors or memory contention that was initially raised as a problem in KIP-844. If this does become a problem, a later KIP could resolve this by tracking the size of the uncommitted records in-memory, and prematurely forcing a commit of the Task when it crosses a configurable threshold. This is outside of the scope of this KIP, as it may not be necessary.

When reading records, we will use the WriteBatchWithIndex#getFromBatchAndDB and WriteBatchWithIndex#newIteratorWithBase utilities in order to ensure that uncommitted writes are available to query. This minimizes the amount of custom code needed to implement transactionality, and allows RocksDB to perform optimizations.

StateStore ownership of checkpointing

One issue with both the existing RocksDBStore and KIP-844 Transactional StateStores that is not resolved by WriteBatchWithIndex is that to guarantee consistency with the Task checkpointing, it is required to explicitly flush the memtables on every commit. This is not what RocksDB was designed for, and leads to sub-optimal performance, especially for lower throughput stores, where many small sstable files are created. This increased "write-amplification" increases pressure on RocksDB's compaction threads, as they have more sstables to compact together.

The existing Task checkpoints are also only written when the Task is cleanly closed. This would cause stores to be wiped and re-built in the event of a crash, even when unnecessary.

To resolve this, we will move the responsibility for StateStore checkpointing to the StateStore interface itself.

When calling StateStore#commit , the offset(s) for the changelog partition, or input partitions for global stores, will be provided to the store. The StateStore itself will then determine how best to checkpoint the data it's committing.

For RocksDBStore, we will store offsets in a separate column-family, offsetMetadata, which is updated as part of the current batch during commit. We ensure that the memtables for our data and metadata column-families are atomically flushed by RocksDB by enabling Atomic Flushes during store initialization. This will guarantee that all records written to the RocksDB memtables, and subsequent on-disk sstables, will always be accompanied by the changelog/input partition offsets that they correspond to, without the need to explicitly flush memtables.

Query Position data

As part of the Intereactive Query v2 (IQv2) initiative, StateStores already track Position offsets, used to bound queries. To improve atomicity, these offsets will also be written to the offsetsMetadata column-family, and atomically committed along with records and changelog/input partition offsets.

Changes to StateManager

Currently, StateManager implementations (ProcessorStateManager for regular stores and GlobalStateManagerImpl for global stores) manage checkpointing of their stores via an on-disk file. This is done through 3 methods in each class:

• flush()
• updateChangelogOffsets(Map)
• checkpoint()

These methods are always called together, with one exception: checkpoint is only called if the number of records processed since the last commit is more than the hard-coded threshold of 10,000. This is presumably a performance optimization to prevent checkpoint files being written every 100 milliseconds.

We will replace these 3 methods, with one:

• commit(Map)

This will delegate the checkpointing procedure to the underlying StateStore.  The StateStoreMetadata , which currently stores the offset for each changelog partition in-memory, will be updated to instead delegate to StateStore#getCommittedOffset().

Compatibility, Deprecation, and Migration Plan

Existing stores maintain their current checkpoints and position offsets in files. These files will still be read, if present, and used to automatically migrate an existing store to being transactional:

For .checkpoint files:

1. If a checkpoint offset file exists with an offset for a changelog partition of an existing store:
   1. The store will be checked for its own offset via StateStore#getCommittedOffset() 
      1. If one exists, the checkpoint file will be ignored.
      2. If one doesn't exist, StateStore#commit(Map) will be called, with the offset(s) from the checkpoint file.
2. If no checkpoint offsets exist in the store, and no checkpoint offset file exists, the store data will be deleted and restored from the changelog.
   • This ensures that corrupt stores will also be properly handled during migration.
   • Only the corrupt store(s) will be deleted. The Task directory itself will not be deleted.
3. The checkpoint offset file will be deleted.

And for .position files:

1. If a position offset file exists for a store:
   1. The store metadata column-family will be checked for existing position offsets.
      1. If none exist, the data in the position file will be written to the metadata column-family.
2. The position offset file will be deleted.

Custom StateStore implementations will continue to operate, however, since Task checkpoints are no longer written, they will be expected to handle their own offsets. If they do not, they will be considered corrupt and wiped. It is up to users to upgrade custom implementations to ensure that this does not happen.

Note: custom implementations that extend an internal implementation, like RocksDBStore, will automatically assume the checkpointing behaviour of that implementation, and should automatically function as expected.

Test Plan

Testing will be accomplished by both the existing tests and by writing some new unit tests that verify atomicity, durability and consistency guarantees that this KIP provides.

Rejected Alternatives

    return false;
    }

   /**      
     * Return an approximate count of memory used by records not yet committed to this StateStore.
     * <p>
     * This method will return an approximation of the memory that would be freed by the next call to {@link
     * #commit(Map)}.
     * <p>
     * If no records have been written to this store since {@link #init(StateStoreContext, StateStore) opening}, or
     * since the last {@link #commit(Map)}; or if this store does not support atomic transactions, it will return {@code
     * 0}, as no records are currently being buffered.
     *
     * @return The approximate size of all records awaiting {@link #commit(Map)}; or {@code 0} if this store does not
     *         support transactions, or has not been written to since {@link #init(StateStoreContext, StateStore)} or
     *         last {@link #commit(Map)}.
     */
    @Evolving
    default long approximateNumUncommittedBytes() {
        return 0;
    }

Metrics

New

• stream-state-metrics 
  • commit-rate - the number of calls to StateStore#commit(Map)
  • commit-latency-avg - the average time taken to call StateStore#commit(Map)
  • commit-latency-max - the maximum time taken to call StateStore#commit(Map)

Deprecated

• stream-state-metrics 
  • flush-rate
  • flush-latency-avg 
  • flush-latency-max 

These changes are necessary to ensure these metrics are not confused with orthogonal operations, like RocksDB memtable flushes or cache flushes. They will be measuring the invocation of StateStore#commit, which replaces StateStore#flush.

While the flush metrics are only deprecated, they will no longer record any data under normal use, as Kafka Streams will no longer call StateStore#flush().

Proposed Changes

To ensure that data is not written to a state store until it has been committed to the changelog, we need to isolate writes from the underlying database until changelog commit. To achieve this, we introduce the concept of transaction Isolation Levels, that dictate the visibility of records, written by processing threads, to Interactive Query threads.

We enable configuration of the level of isolation provided by StateStores via a default.state.isolation.level, which can be configured to either:

In Kafka Streams, all StateStore s are written to by a single StreamThread  (this is the Single Writer principle). However, multiple other threads may concurrently read from StateStore s, principally to service Interactive Queries. In practice, this means that under READ_COMMITTED, writes by the StreamThread  that owns the StateStore  will only become visible to Interactive Query threads once commit()  has been called.

The default value for default.state.isolation.level will be READ_UNCOMMITTED, to mirror the behaviour we have today; but this will be automatically set to READ_COMMITTED if the processing.mode has been set to an EOS mode, and the user has not explicitly set deafult.state.isolation.level to READ_UNCOMMITTED. This will provide EOS users with the most useful behaviour out-of-the-box, but ensures that they may choose to sacrifice the benefits of transactionality to ensure that Interactive Queries can read records before they are committed, which is required by a minority of use-cases.

In-memory Transaction Buffers

Many StateStore implementations, including RocksDB, will buffer records written to a transaction entirely in-memory, which could cause issues, either with JVM heap or native memory. To mitigate this, we will automatically force a Task commit if the total memory used for buffering uncommitted records returned by StateStore#approximateNumUncommittedBytes() exceeds the threshold configured by statestore.uncommitted.max.bytes. This will roughly bound the memory required for buffering uncommitted records, irrespective of the commit.interval.ms, and will effectively bound the number of records that will need to be restored in the event of a failure. Each StreamThread will be given 1/num.stream.threads of the configured limits, dividing it fairly between them.

It's possible that some Topologies can generate many more new StateStore entries than the records they process, in which case, it would be possible for such a Topology to cross the configured record/memory thresholds mid-processing, potentially causing an OOM error if these thresholds are exceeded by a lot. To mitigate this, the StreamThread will measure the increase in records/bytes written on each iteration, and pre-emptively commit if the next iteration is likely to cross the threshold.

Note that this new method provides default implementations that ensure existing custom stores and non-transactional stores (e.g. InMemoryKeyValueStore) do not force any early commits.

Interactive Queries

Interactive queries currently see every record, as soon as they are written to a StateStore. This can cause some consistency issues, as interactive queries can read records before they're committed to the Kafka changelog, which may be rolled-back. To address this, we have introduced configurable isolation levels, configured globally via default.state.isolation.level (see above).

When operating under the READ_COMMITTED isolation level, the maximum time for records to become visible to interactive queries will be commit.interval.ms. Under EOS, this is by default a low value (100 ms), but under at-least-once, the default is 30 seconds. Users may need to adjust their commit.interval.ms to meet the visibility latency goals for their use-case.

When operating under the READ_UNCOMMITTED isolation level, (i.e. ALOS), all records will be immediately visible to interactive queries, so the high default commit.interval.ms of 30s will have no impact on interactive query latency.

Error Handling

Kafka Streams currently generates a TaskCorruptedException when a Task needs to have its state wiped (under EOS) and be re-initialized. There are currently several different situations that generate this exception:

1. No offsets for the store can be found when opening it under EOS.
2. OutOfRangeException during restoration, usually caused by the changelog being wiped on application reset.
3. TimeoutException under EOS, when writing to or committing a Kafka transaction.

The first two of these are extremely rare, and make sense to keep. However, timeouts are much more frequent. They currently require the store to be wiped under EOS because when a timeout occurs, the data in the local StateStore will have been written, but the data in the Kafka changelog will have failed to be written, causing a mismatch in consistency.

With Transactional StateStores, we can guarantee that the local state is consistent with the changelog, therefore, it will no longer be necessary to reset the local state on a TimeoutException when operating under the READ_COMMITTED isolation level.

Atomic Checkpointing

Kafka Streams currently stores the changelog offsets for a StateStore in a per-Task on-disk file, .checkpoint, which under EOS, is written only when Streams shuts down successfully. There are two major problems with this approach:

• To ensure that the data on-disk matches the checkpoint offsets in the .checkpoint file, we must flush the StateStores whenever we update the offsets in .checkpoint. This is a performance regression, as it causes a significant increase in the frequency of RocksDB memtable flushes, which increases load on RocksDB's compaction threads.
• There's a race condition, where it's possible the application exits after data has been committed to RocksDB, but before the checkpoint file has been updated, causing a consistency violation.

To resolve this, we move the responsibility for offset management to the StateStore itself. The new commit method takes a map of all the changelog offsets that correspond to the state of the transaction buffer being committed.

RocksDBStore will store these offsets in a separate Column Family, and will be configured to atomically flush all its Column Families. This guarantees that the changelog offsets will always be flushed to disk together with the data they represent, irrespective of how that flush is triggered. This allows us to remove the explicit memtable flush(), enabling RocksDB to dictate when memtables are flushed to disk.

The existing .checkpoint files will be retained for any StateStore that does not set managesOffsets()  to true , and to ensure managed offsets are available when the store is closed. Existing offsets will be automatically migrated into StateStores that manage their own offsets, iff there is no offset returned by StateStore#committedOffset.

Required interface changes:

• Add methods void commit(Map<TopicPartition, Long> changelogOffsets), boolean managesOffsets() and Long committedOffset(TopicPartition) to StateStore .
• Deprecate method flush() on StateStore.

Offsets for Consumer Rebalances

Kafka Streams directly reads from the Task .checkpoint file during Consumer rebalance, in order to optimize assignments of stateful Tasks by assigning them to the instance with the most up-to-date copy of the data, which minimises restoration. To allow this to continue functioning, Kafka Streams will continue to write the changelog offsets to the .checkpoint file, even for stores that manage their own offsets.

Offsets will be written to .checkpoint at the following times:

1. During StateStore initialization, in order to synchronize the offsets in .checkpoint  with the offsets returned by StateStore#committedOffset(TopicPartition), which are the source of truth for stores that manage their own offsets.
2. When the StateStore is closed, in order to ensure that the offsets used for Task assignment reflect the state persisted to disk.
3. At the end of every Task commit, if-and-only-if at least one StateStore in the Task is persistent and does not manage its own offsets. This ensures that stores that don't manage their offsets continue to have their offsets persisted to disk whenever the StateStore data itself is committed.
   • Avoiding writing .checkpoint when every persistent store manages its own offsets ensures we don't pay a significant performance penalty when the commit interval is short, as it is by default under EOS.
   • Since all persistent StateStores provided by Kafka Streams will manage their own offsets, the common case is that the .checkpoint file will not be updated on commit(Map) 

Tasks that are already assigned to an instance, already use the in-memory offsets when calculating partition assignments, so no change is necessary here.

Interactive Query .position Offsets

Input partition "Position" offsets, introduced by KIP-796: Interactive Query v2, are currently stored in a .position file by the RocksDBStore implementation. To ensure consistency with the committed data and changelog offsets, these position offsets will be stored in RocksDB, in the same column family as the changelog offsets, instead of the .position file. When a StateStore that manages its own offsets is first initialized, if a .position file exists in the store directory, its offsets will be automatically migrated into the store, and the file will be deleted.

When writing data to a RocksDBStore (via put, delete, etc.), the input partition offsets will be read from the changelog record metadata (as before), and these offsets will be added to the current transactions WriteBatch. When the StateStore is committed, the position offsets in the current WriteBatch will be written to RocksDB, alongside the records they correspond to. Alongside this, RocksDBStore will maintain two Position maps in-memory, one containing the offsets pending in the current transaction's WriteBatch, and the other containing committed offsets. On commit(Map), the uncommitted Position map will be merged into the committed Position map. In this sense, the two Position maps will diverge during writes, and re-converge on-commit.

When an interactive query is made under the READ_COMMITTED isolation level the PositionBound will constrain the committed Position map, whereas under READ_UNCOMMITTED, the PositionBound will constrain the uncommitted Position map.

RocksDB Transactions

When the isolation level is READ_COMMITTED, we will use RocksDB's WriteBatchWithIndex as a means to accomplishing atomic writes when not using the RocksDB WAL. When reading records from the StreamThread, we will use the WriteBatchWithIndex#getFromBatchAndDB and WriteBatchWithIndex#newIteratorWithBase utilities in order to ensure that uncommitted writes are available to query. When reading records from Interactive Queries, we will use the regular RocksDB#get and RocksDB#newIterator methods, to ensure we see only records that have been committed (see above). The performance of this is expected to actually be better than the existing, non-batched write path. The main performance concern is that the WriteBatch must reside completely in-memory until it is committed, which is addressed by statestore.uncommitted.max.bytes, see above.

Compatibility, Deprecation, and Migration Plan

The above changes will retain compatibility for all existing StateStores, including user-defined custom implementations. Any StateStore that extends RocksDBStore will automatically inherit its behaviour, although its internals will change, potentially requiring users that depend on internal behaviour to update their code.

All new methods on existing classes will have defaults set to ensure compatibility.

Kafka Streams will automatically migrate offsets found in an existing .checkpoint file, and/or an existing .position file, to store those offsets directly in the StateStore, if managesOffsets returns true. Users of the in-built store types will not need to make any changes. See Upgrading.

Users may notice a change in the performance/behaviour of Kafka Streams. Most notably, under EOS Kafka Streams will now regularly "commit" StateStores, where it would have only done so when the store was closing in the past. The overall performance of this should be at least as good as before, but the profile will be different, with write latency being substantially faster, and commit latency being a bit higher.

Upgrading

When upgrading to a version of Kafka Streams that includes the changes outlined in this KIP, users will not be required to take any action. Kafka Streams will automatically upgrade any RocksDB stores to manage offsets directly in the RocksDB database, by importing the offsets from any existing .checkpoint and/or .position files.

Users that currently use processing.mode: exactly-once(-v2|-beta) and who wish to continue to read uncommitted records from their Interactive Queries will need to explicitly set default.state.isolation.level: READ_UNCOMMITTED.

Downgrading

When downgrading from a version of Kafka Streams that includes the changes outlined in this KIP to a version that does not contain these changes, users will not be required to take any action. The older Kafka Streams version will be unable to open any RocksDB stores that were upgraded to store offsets (see Upgrading), which will cause Kafka Streams to wipe the state for those Tasks and restore the state, using an older RocksDB store format, from the changelogs.

Since downgrading is a low frequency event, and since restoring state from scratch is already an existing failure mode for older versions of Kafka Streams, we deem this an acceptable automatic downgrade strategy.

Test Plan

Testing will be accomplished by both the existing tests and by writing some new unit tests that verify atomicity, durability and consistency guarantees that this KIP provides.

Rejected Alternatives

Dual-Store Approach (KIP-844)

The design outlined in KIP-844, sadly, does not perform well (as described above), and requires users to opt-in to transactionality, instead of being a guarantee provided out-of-the-box.

Replacing RocksDB memtables with ThreadCache

It was pointed out on the mailing list that Kafka Streams fronts all RocksDB StateStores with a configurable record cache, and that this cache duplicates the function requests for recently written records provided by RocksDB memtables. A suggestion was made to utilize this record cache (the ThreadCache class) as a replacement for memtables, by directly flushing them to SSTables using the RocksDB SstFileWriter.

This is out of scope of this KIP, as its goal would be reducing the duplication (and hence, memory usage) of RocksDB StateStores; whereas this KIP is tasked with improving the consistency of StateStores to reduce the frequency and impact of state restoration, improving their scalability.

It has been recommended to instead pursue this idea in a subsequent KIP, as the interface changes outlined in this KIP should be compatible with this idea.

Transactional support under READ_UNCOMMITTED

When query isolation level is READ_UNCOMMITTED, Interactive Query threads need to read records from the ongoing transaction buffer. Unfortunately, the RocksDB WriteBatch is not thread-safe, causing Iterators created by Interactive Query threads to produce invalid results/throw unexpected errors as the WriteBatch is modified/closed during iteration.

Ideally, we would build an implementation of a transaction buffer that is thread-safe, enabling Interactive Query threads to query it safely. One approach would be to "chain together" WriteBatches, creating a new WriteBatch every time a new Iterator is created by an Interactive Query thread and "freezing" the previous WriteBatch.

It was decided to defer tackling this problem to a later KIP, in order to realise the benefits of transactional state stores to users as quickly as possible.

Query-time Isolation Levels

It was requested that users be able to select the isolation level of queries on a per-query basis. This would require some additional API changes (to the Interactive Query APIs). Such an API would require that state stores are always transactional, and that the transaction buffers can be read from by READ_UNCOMMITTED queries. Due to the problems outlined in the previous section, it was decided to also defer this to a subsequent KIP.

The new configuration option default.state.isolation.level was deliberately named to enable query-time isolation levels in the future, whereby any query that didn't explicitly choose an isolation level would use the configured default. Until then, this configuration option will globally control the isolation level of all queries, with no way to override it per-queryThe design outlined in KIP-844, sadly, does not perform well (as described above), and requires users to opt-in to transactionality, instead of being a guarantee provided out-of-the-box.