[
https://issues.apache.org/jira/browse/KAFKA-2260?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14632113#comment-14632113
]
Ben Kirwin commented on KAFKA-2260:
-----------------------------------
Ah, sorry! Let me try again.
Suppose you try and send a batch of messages to a partition, but get some
network error -- it's possible that they were published successfully, or that
they were lost before they made it to the broker. With the CAS, the simple
thing to do is to resend the same batch with the same expected offsets. If the
messages were published correctly last time, you'll get a check mismatch error;
and if they weren't, they'll be appended correctly.
If the series of messages that the producer wants to send is fixed, the same
mechanism would work even through producer restarts. If the set of messages
isn't fixed -- the producer might have a completely different set of messages
to send after restarting -- than what it means to be exactly-once becomes a lot
more domain-dependent; you might want to write exactly one group of messages
for each input message, or rpc request, or five-minute interval -- but that
requires coordination between a bunch of different moving parts, and I don't
think there's one coordination mechanism that handles all cases. (This
'expected offset' thing is enough for some, but certainly not all of them...)
> Allow specifying expected offset on produce
> -------------------------------------------
>
> Key: KAFKA-2260
> URL: https://issues.apache.org/jira/browse/KAFKA-2260
> Project: Kafka
> Issue Type: Improvement
> Reporter: Ben Kirwin
> Assignee: Ewen Cheslack-Postava
> Priority: Minor
> Attachments: expected-offsets.patch
>
>
> I'd like to propose a change that adds a simple CAS-like mechanism to the
> Kafka producer. This update has a small footprint, but enables a bunch of
> interesting uses in stream processing or as a commit log for process state.
> h4. Proposed Change
> In short:
> - Allow the user to attach a specific offset to each message produced.
> - The server assigns offsets to messages in the usual way. However, if the
> expected offset doesn't match the actual offset, the server should fail the
> produce request instead of completing the write.
> This is a form of optimistic concurrency control, like the ubiquitous
> check-and-set -- but instead of checking the current value of some state, it
> checks the current offset of the log.
> h4. Motivation
> Much like check-and-set, this feature is only useful when there's very low
> contention. Happily, when Kafka is used as a commit log or as a
> stream-processing transport, it's common to have just one producer (or a
> small number) for a given partition -- and in many of these cases, predicting
> offsets turns out to be quite useful.
> - We get the same benefits as the 'idempotent producer' proposal: a producer
> can retry a write indefinitely and be sure that at most one of those attempts
> will succeed; and if two producers accidentally write to the end of the
> partition at once, we can be certain that at least one of them will fail.
> - It's possible to 'bulk load' Kafka this way -- you can write a list of n
> messages consecutively to a partition, even if the list is much larger than
> the buffer size or the producer has to be restarted.
> - If a process is using Kafka as a commit log -- reading from a partition to
> bootstrap, then writing any updates to that same partition -- it can be sure
> that it's seen all of the messages in that partition at the moment it does
> its first (successful) write.
> There's a bunch of other similar use-cases here, but they all have roughly
> the same flavour.
> h4. Implementation
> The major advantage of this proposal over other suggested transaction /
> idempotency mechanisms is its minimality: it gives the 'obvious' meaning to a
> currently-unused field, adds no new APIs, and requires very little new code
> or additional work from the server.
> - Produced messages already carry an offset field, which is currently ignored
> by the server. This field could be used for the 'expected offset', with a
> sigil value for the current behaviour. (-1 is a natural choice, since it's
> already used to mean 'next available offset'.)
> - We'd need a new error and error code for a 'CAS failure'.
> - The server assigns offsets to produced messages in
> {{ByteBufferMessageSet.validateMessagesAndAssignOffsets}}. After this
> changed, this method would assign offsets in the same way -- but if they
> don't match the offset in the message, we'd return an error instead of
> completing the write.
> - To avoid breaking existing clients, this behaviour would need to live
> behind some config flag. (Possibly global, but probably more useful
> per-topic?)
> I understand all this is unsolicited and possibly strange: happy to answer
> questions, and if this seems interesting, I'd be glad to flesh this out into
> a full KIP or patch. (And apologies if this is the wrong venue for this sort
> of thing!)
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
