[v22.1.x] Improve idempotency latency by introducing rm_stm pipelining #5450
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
rystsov
requested review from
dotnwat,
NyaliaLui,
BenPope,
mmaslankaprv,
ztlpn,
jcsp and
VadimPlh
as code owners
fixes redpanda-data#5189 (cherry picked from commit 41e7d2f)
(cherry picked from commit 93b8bd4)
(cherry picked from commit 9a693fc)
(cherry picked from commit 0791a5f)
(cherry picked from commit 9b5491b)
Return invalid_request instead of processing a wrong if branch (cherry picked from commit 8ea6b14)
Kafka protocol is asynchronous: a client may send the next write
request without waiting for the previous request to finish. Since
Redpanda's Raft implementation is also asynchronous, it's possible
to process Kafka requests with minimal overhead and/or delays.
The challenge is to process causally related requests without
pipeline stalling. Imagine that we have two requests A and B. When
we want to process them as fast as possible we should start repli-
cating the latter request without waiting until the replication of
the first request is over.
But in case the requests are the RSM commands and are causally
related e.g. request A is "set x=2" and B is "set y=3 if x=2" then
to guarantee that B's precondition holds Redpanda can't start
replication of B request without knowing that
- The request A is successfully executed
- No other command sneaks in between A and B
In case of idempotency the condition we need to preserve is the
monotonicity of the seq numbers without gaps.
In order to preserve causality and to avoid pipeline stalls we use
optimistic replication. A leader knows about its ongoing operations
so it may optimistically predict the outcome of operation A and
start executing operation B assuming its prediction is true.
But what happens if Redpanda is wrong with its prediction, how does
it ensure safety? Let's break down all the failure scenarios.
A leader uses a mutex and consesnsus::replicate_in_stages order all
the incoming requests. Before the first stage is resolves the mutex
controls the order and after the first stage is resolved Redpanda's
raft implementation guarantees order:
auto u = co_await mutex.lock();
auto stages = raft.replicate_in_stages(request);
co_await stages.enqueued;
u.return_all();
// after this point the order between the requests is certain
// the order enforced by mutex is preserved in the log
The mutex / replicate_in_stages combination may enforce partial
order but can't prevent out-of-nowhere writes. Imagine a node loses
the leadership then new leader inserts a command C and transfers
the leadership back to the original node.
To fight this Redpanda uses conditional replication:
auto term = persisted_stm::sync();
auto u = co_await mutex.lock();
auto stages = raft.replicate_in_stages(term, request);
co_await stages.enqueued;
u.return_all();
It uses a sync method to make sure that the RSM's state reflects
all the commands replicated by previous term (during this phase it
may learn about the "A" command) and then it uses the raft's term
to issue the replication command (conditional replicate). By design
the replication call can't succeed if the term is wrong so instead
of leading the ACB state the replication of B is doomed to fail.
Redpanda's Raft implementation guarantees that if A was enqueued
before B within the same term then the replication of B can't be
successful without replication of A so we may not worry that A may
be dropped.
But we still have uncertainty. Should Redpanda process B assuming
that A was successful or should it assume it failed? In order to
resolve the uncertainty the leader steps down.
Since sync() guarantees that the RSM's state reflects all the
commands replicated by previous term - the next leader will resolve
the uncertainty.
The combination of those methods lead to preserving causal
relationships between the requests (idempotency) without
introducing pipelining stalls.
-------------------------------------------------------------------
The idempotency logic is straightforward:
- Take a lock identified by producer id (concurrent producers
don't affect each other)
- If the seq number and its offset is already known then
return the offset
- If the seq number is known and in flight then "park" the
current request (once the original resolved it pings all
the parked requests with the written offsets or an error)
- If the seq number is seen for the first time then:
- Reject it if there is a gap between it and the last
known seq
- Start processing if there is no gap
fixes redpanda-data#5054
(cherry picked from commit 88423d7)
Leadership transfer and rm_stm should be coordinated to avoid a possibility of the live lock: - leadership transfer sets _transferring_leadership - it causes a write to fail - failing write initiate leader step down - the step down causes the transfer to fail (cherry picked from commit 68848c1)
(cherry picked from commit 1064521)
making transfer_leadership_to more stable by checking that all nodes has updated metadata. previously we could choose a stale server, get 503 and retry the requests (cherry picked from commit f15e3d8)
(cherry picked from commit 9ccbfce)
temporary adding a faster version of wait_until before the ducktape repo is updated (cherry picked from commit b4a5fb2)
bharathv
approved these changes
Jul 13, 2022
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Backport from pull request #5157, #5204
fixes #5445