Change history for Faust 1.0

This document contain change notes for bugfix releases in the Faust 1.x series. If you’re looking for changes in the latest series, please visit the latest Changelog.

For even older releases you can visit the History section.

1.0.30

release-date:

2018-08-15 3:17 P.M PDT

release-by:

Ask Solem

  • Requirements

  • Typing: faust.types.Message.timestamp_type is now the correct

    int, previously it was string by message.

  • Models: Records can now have recursive fields.

    For example a tree structure model having a field that refers back to itself:

    class Node(faust.Record):
    data: Any
    children: List['Node']
  • Models: A field of type List[Model] no longer raises an exception

    if the value provided is None.

  • Models: Adds support for --strict-optional-style fields.

    Previously the following would work:

    class Order(Record):
    account: Account = None

    The account is considered optional from a typing point of view, but only if the https://pypi.org/project/mypy/ option --strict-optional is disabled.

    Now that --strict-optional is enabled by default in https://pypi.org/project/mypy/, this version adds support for fields such as:

    class Order(Record):
    account: Optional[Account] = None
    history: Optional[List[OrderStatus]]
  • Models: Class options such as isodates/include_metadata/etc. are

    now inherited from parent class.

  • Stream: Fixed NameError when pushing non-Event value into stream.

1.0.29

release-date:

2018-08-10 5:00 P.M PDT

release-by:

Vineet Goel

  • Requirements

  • Table: Fixed hanging at startup/rebalance on Python 3.7 (Issue #134).

    Workaround for asyncio bug seemingly introduced in Python 3.7, that left the worker hanging at startup when attempting to recover a table without any data.

  • Monitor: More efficient updating of highwater metrics (Issue #139).

  • Partition Assignor: The assignor now compresses the metadata being passed around to all application instances for efficiency and to avoid extreme cases where the metadata is too big.

1.0.28

release-date:

2018-08-08 11:25 P.M PDT

release-by:

Vineet Goel

  • Monitor: Adds consumer stats such as last read offsets, last committed offsets and log end offsets to the monitor. Also added to the StatsdMonitor.

  • aiokafka: Changes how topics are created to make it more efficient. We now are smarter about finding kafka cluster controller instead of trial and error.

  • Documentation: Fixed links to Slack and other minor fixes.

1.0.27

release-date:

2018-07-30 04:00 P.M PDT

release-by:

Ask Solem

  • No code changes

  • Fixed links to documentation in README.rst

1.0.26

release-date:

2018-07-30 08:00 A.M PDT

release-by:

Ask Solem

  • Public release.

1.0.25

release-date:

2018-07-27 12:43 P.M PDT

release-by:

Ask Solem

  • stream_publish_on_commit accidentally disabled by default.

    This made the rate of producing much slower, as the default buffering settings are not optimized.

  • The App.rebalancing flag is now reset after the tables have recovered.

1.0.24

release-date:

2018-07-12 6:54 P.M PDT

release-by:

Ask Solem

1.0.23

release-date:

2018-07-11 5:00 P.M PDT

release-by:

Ask Solem

  • Requirements

  • Now compatible with Python 3.7.

  • Setting stream_wait_empty is now disabled by default (Issue #117).

  • Documentation build now compatible with Python 3.7.

    • Fixed ForwardRef has no attribute __origin__ error.

    • Fixed DeprecatedInSphinx2.0 warnings.

  • Web: Adds app.on_webserver_init(web) callback for ability to serve static files using web.add_static.

  • Web: Adds web.add_static(prefix, fs_path)

  • Worker: New App.unassigned attribute is now set if the worker does not have any assigned partitions.

  • CLI: Console colors was disabled by default.

1.0.22

release-date:

2018-06-27 5:35 P.M PDT

release-by:

Vineet Goel

  • aiokafka: Timeout for topic creation now wraps entire topic creation. Earlier this timeout was for each individual request.

  • testing: Added stress testing suite.

1.0.21

release-date:

2018-06-27 1:43 P.M PDT

release-by:

Ask Solem

Warning

This changes the package name of kafka to rhkafka.

1.0.20

release-date:

2018-06-26 2:35 P.M PDT

release-by:

Vineet Goel

  • Monitor: Added Monitor.count to add arbitrary metrics to app monitor.

  • Statsd Monitor: Normalize agent metrics by removing memory address to avoid spamming statsd with thousands of unique metrics per agent.

1.0.19

release-date:

2018-06-25 6:40 P.M PDT

release-by:

Vineet Goel

  • Assignor: Fixed crash if initial state of assignment is invalid. This was causing the following error: ValueError('Actives and Standbys are disjoint',). during partition assignment.

1.0.18

release-date:

2018-06-21 3:53 P.M PDT

release-by:

Ask Solem

  • Worker: Fixed KeyError: TopicPartition(topic='...', partition=x) occurring during rebalance.

1.0.17

release-date:

2018-06-21 3:15 P.M PDT

release-by:

Ask Solem

1.0.16

release-date:

2018-06-19 3:46 P.M PDT

release-by:

Ask Solem

  • Worker: https://pypi.org/project/aiokafka/ create topic request default timeout now set

    to 20 seconds (previously it was accidentally set to 1000 seconds).

  • Worker: Fixes crash from AssertionError where table._revivers

    is an empty list.

  • Distribution: Adds t/misc/scripts/rebalance/killer-always-same-node.sh.

1.0.15

release-date:

2018-06-14 7:36 P.M PDT

release-by:

Ask Solem

  • Requirements

  • Worker: Fixed problem where worker does not recover after MacBook sleeping and waking up.

  • Worker: Fixed crash that could lead to rebalancing loop.

  • Worker: Removed some noisy errors that weren’t really errors.

1.0.14

release-date:

2018-06-13 5:58 P.M PDT

release-by:

Ask Solem

  • Requirements

  • Worker: https://pypi.org/project/aiokafka/’s heartbeat thread would sometimes keep the worker alive even though the worker was trying to shutdown.

    An error could have happened many hours ago causing the worker to crash and attempt a shutdown, but then the heartbeat thread kept the worker from terminating.

    Now the rebalance will check if the worker is stopped and then appropriately stop the heartbeat thread.

  • Worker: Fixed error that caused rebalancing to hang: "ValueError: Set of coroutines/Futures is empty.".

  • Worker: Fixed error “Coroutine x tried to break fence owned by y”

    This was added as an assertion to see if multiple threads would use the variable at the same time.

  • Worker: Removed logged error “not assigned to topics” now that we automatically recover from non-existing topics.

  • Tables: Ignore asyncio.CancelledError while stopping standbys.

  • Distribution: Added scripts to help stress test rebalancing in t/misc/scripts/rebalance.

1.0.13

release-date:

2018-06-12 2:10 P.M PDT

release-by:

Ask Solem

  • Worker: The Kafka fetcher service was taking too long to shutdown on rebalance.

    If this takes longer than the session timeout, it triggers another rebalance, and if it happens repeatedly this will cause the cluster to be in a state of constant rebalancing.

    Now we use future cancellation to stop the service as fast as possible.

  • Worker: Fetcher was accidentally started too early.

    This didn’t lead to any problems that we know of, but made the start a bit slower than it needs to.

  • Worker: Fixed race condition where partitions were paused while fetching from them.

  • Worker: Fixed theoretical race condition hang if web server started and stopped in quick succession.

  • Statsd: The statsd monitor prematurely initialized the event loop on module import.

    We had a fix for this, but somehow forgot to remove the “hard coded super” that was set to call: Service.__init__(self, **kwargs).

    The class is not even a subclass of Service anymore, and we are lucky it manifests merely when doing something drastic, like py.test, recursively importing all modules in a directory.

1.0.12

release-date:

2018-06-06 1:34 P.M PDT

release-by:

Ask Solem

  • Requirements

  • Worker: Producer crashing no longer causes the consumer to hang at shutdown while trying to publish attached messages.

1.0.11

release-date:

2018-05-31 16:41 P.M PDT

release-by:

Ask Solem

1.0.10

release-date:

2018-05-15 16:02 P.M PDT

release-by:

Vineet Goel

  • Worker: Stop reading changelog when no remaining messages.

1.0.9

release-date:

2018-05-15 15:42 P.M PDT

release-by:

Vineet Goel

  • Worker: Do not stop reading standby updates.

1.0.8

release-date:

2018-05-15 11:00 A.M PDT

release-by:

Vineet Goel

  • Tables

    • Fixes bug due to which we were serializing None values while recording a key delete to the changelog. This was causing the deleted keys to never be deleted from the changelog.

    • We were earlier not persisting offsets of messages read during changelog reading (or standby recovery). This would cause longer recovery times if recovery was ever interrupted.

  • App: Added flight recorder for consumer group rebalances for debugging.

1.0.7

release-date:

2018-05-14 4:53 P.M PDT

release-by:

Ask Solem

  • Requirements

  • App: key_type and value_type can now be set to:

    • int: key/value is number stored as string

    • float: key/value is floating point number stored as string.

    • decimal.Decimal key/value is decimal stored as string.

  • Agent: Fixed support for group_by/through after change to reuse the same stream after agent crashing.

  • Agent: Fixed isolated_partitions=True after change in v1.0.3.

    Initialization of the agent-by-topic index was in 1.0.3 moved to the AgentManager.start method, but it turns out AgentManager is a regular class, and not a service.

    AgentManager is now a service responsible for starting/stopping the agents required by the app.

  • Agent: Include active partitions in repr when isolated_partitions=True.

  • Agent: Removed extraneous ‘agent crashed’ exception in logs.

  • CLI: Fixed autodiscovery of commands when using faust -A app.

  • Consumer: Appropriately handle closed fetcher.

  • New shortcut: faust.uuid() generates UUID4 ids as string.

1.0.6

release-date:

2018-05-11 11:15 A.M PDT

release-by:

Vineet Goel

  • Requirements:

    • Now depends on Aiokafka 0.4.7.

  • Table: Delete keys when raw value in changelog set to None

    This was resulting in deleted keys still being present with value None upon recovery.

  • Transports: Crash app on CommitFailedError thrown by https://pypi.org/project/aiokafka/.

    App would get into a weird state upon a commit failed error thrown by the consumer thread in the https://pypi.org/project/aiokafka/ driver.

1.0.5

release-date:

2018-05-08 4:09 P.M PDT

release-by:

Ask Solem

  • Requirements:

  • Agents: Fixed problem with hanging after agent raises exception.

    If an agent raises an exception we cannot handle it within the stream iteration, so we need to restart the agent.

    Starting from this change, even though we restart the agent, we reuse the same faust.Stream object that the crashed agent was using.

    This makes recovery more seamless and there are fewer steps involved.

  • Transports: Fixed worker hanging issue introduced in 1.0.4.

    In version 1.0.4 we introduced a bug in the round-robin scheduling of topic partitions that manifested itself by hanging with 100% CPU usage.

    After processing all records in all topic partitions, the worker would spin loop.

  • API: Added new base class for windows: faust.Window

    There was the typing interface faust.types.windows.WindowT, but now there is also a concrete base class that can be used in for example Mock(autospec=Window).

  • Tests: Now takes advantage of the new AsyncMock.

1.0.4

release-date:

2018-05-08 11:45 A.M PDT

release-by:

Vineet Goel

  • Transports:

    In version-1.0.2 we implemented fair scheduling in https://pypi.org/project/aiokafka/ transport such that while processing the worker had an equal chance of processing each assigned Topic. Now we also round-robin through topic partitions within topics such that the worker has an equal chance of processing message from each assigned partition within a topic as well.

1.0.3

release-date:

2018-05-07 3:45 P.M PDT

release-by:

Ask Solem

  • Tests:

    • Adds 5650 lines of tests, increasing test coverage to 90%.

  • Requirements:

  • Development:

    • CI now builds coverage.

    • CI now tests multiple CPython versions:

      • CPython 3.6.0

      • CPython 3.6.1

      • CPython 3.6.2

      • CPython 3.6.3

      • CPython 3.6.4

      • CPython 3.6.5

  • Backward incompatible changes:

    • Removed faust.Set unused by any internal applications.

  • Fixes:

    • app.agents did not forward app to AgentManager.

      The agent manager does not use the app, but fixing this in anticipation of people writing custom agent managers.

    • AgentManager: On partitions revoked

      the agent manager now makes sure there’s only one call to each agents agent.on_partitions_revoked callback.

      This is more of a pedantic change, but could have caused problems for advanced topic configurations.

1.0.2

release-date:

2018-05-03 3:32 P.M PDT

release-by:

Ask Solem

  • Transports: Implements fair scheduling in https://pypi.org/project/aiokafka/ transport.

    We now round-robin through topics when processing fetched records from Kafka. This helps us avoid starvation when some topics have many more records than others, and also takes into account that different topics may have wildly varying partition counts.

    In this version when a worker is subscribed to partitions:

    [
    TP(topic='foo', partition=0),
    TP(topic='foo', partition=1),
    TP(topic='foo', partition=2),
    TP(topic='foo', partition=3),

    TP(topic='bar', partition=0),
    TP(topic='bar', partition=1),
    TP(topic='bar', partition=2),
    TP(topic='bar', partition=3),

    TP(topic='baz', partition=0)
]

    Note

    TP is short for topic and partition.

    When processing messages in these partitions, the worker will round robin between the topics in such a way that each topic will have an equal chance of being processed.

  • Transports: Fixed crash in https://pypi.org/project/aiokafka/ transport.

    The worker would attempt to commit an empty set of partitions, causing an exception to be raised. This has now been fixed.

  • Stream: Removed unused method Stream.tee.

    This method was an example implementation and not used by any of our internal apps.

  • Stream: Fixed bug when something raises StopAsyncIteration

    while processing the stream.

    The Python async iterator protocol mandates that it’s illegal to raise StopAsyncIteration in an __aiter__ method.

    Before this change, code such as this:

    async for value in stream:
    value = anext(other_async_iterator)

    where anext raises StopAsyncIteration, Python would have the outer __aiter__ reraise that exception as:

    RuntimeError('__aiter__ raised StopAsyncIteration')

    This no longer happens as we catch the StopAsyncIteration exception early to ensure it does not propagate.

1.0.1

release-date:

2018-05-01 9:52 A.M PDT

release-by:

Ask Solem

  • Stream: Fixed issue with using break when iterating over stream.

    The last message in a stream would not be acked if the break keyword was used:

    async for value in stream:
    if value == 3:
        break

  • Stream: .take now acks events after buffer processed.

    Previously the events were erroneously acked at the time of entering the buffer.

    Note

    To accomplish this we maintain a list of events to ack as soon as the buffer is processed. The operation is O(n) where n is the size of the buffer, so please keep buffer sizes small (e.g. 1000).

    A large buffer will increase the chance of consistency issues where events are processed more than once.

  • Stream: New noack modifier disables acking of messages in the stream.

    Use this to disable automatic acknowledgment of events:

    async for value in stream.noack():
    # manual acknowledgment
    await stream.ack(stream.current_event)

    Manual Acknowledgment

    The stream is a sequence of events, where each event has a sequence number: the “offset”.

    To mark an event as processed, so that we do not process it again, the Kafka broker will keep track of the last committed offset for any topic.

    This means “acknowledgment” works quite differently from other message brokers, such as RabbitMQ where you can selectively ack some messages, but not others.

    If the messages in the topic look like this sequence:

    1 2 3 4 5 6 7 8

    You can commit the offset for #5, only after processing all events before it. This means you MUST ack offsets (1, 2, 3, 4) before being allowed to commit 5 as the new offset.

  • Stream: Fixed issue with .take not properly respecting the within argument.

    The new implementation of take now starts a background thread to fill the buffer. This avoids having to restart iterating over the stream, which caused issues.

1.0.0

release-date:

2018-04-27 4:13 P.M PDT

release-by:

Ask Solem

  • Models: Raise error if Record.asdict() is overridden.

  • Models: Can now override Record._prepare_dict to change the payload generated.

    For example if you want your model to serialize to a dictionary, but not have any fields with None values, you can override _prepare_dict to accomplish this:

    class Quote(faust.Record):
    ask_price: float = None
    bid_price: float = None

    def _prepare_dict(self, data):
        # Remove keys with None values from payload.
        return {k: v for k, v in data.items() if v is not None}

assert Quote(1.0, None).asdict() == {'ask_price': 1.0}

  • Stream: Removed annoying Flight Recorder logging that was too noisy.