796c04e7a8
test: fix flaky test
...
We should emit the trace event before replying to callers.
Example failure:
https://github.com/emqx/emqx/actions/runs/8378977952/job/22946318696#step:6:182
```
=CRITICAL REPORT==== 21-Mar-2024::17:45:37.676024 ===
"check stage" failed: error
{assertMatch,[{module,emqx_ds_storage_bitfield_lts_SUITE},
{line,270},
{expression,"? of_kind ( emqx_ds_replication_layer_egress_flush , Trace )"},
{pattern,"[ # { batch := [ _ , _ , _ ] } ]"},
{value,[]}]}
Stacktrace: [{emqx_ds_storage_bitfield_lts_SUITE,
'-t_atomic_store_batch/1-fun-1-',1,
[{file,
"/__w/emqx/emqx/apps/emqx_durable_storage/test/emqx_ds_storage_bitfield_lts_SUITE.erl"},
{line,270}]},
{emqx_ds_storage_bitfield_lts_SUITE,t_atomic_store_batch,1,
[{file,
"/__w/emqx/emqx/apps/emqx_durable_storage/test/emqx_ds_storage_bitfield_lts_SUITE.erl"},
{line,249}]}]
```
2024-03-21 15:47:29 -03:00
68af211130
fix(ds): reply sync callers after raft store failure
2024-03-21 15:40:21 -03:00
70737a437a
fix(ds): add caller to pending replies before flushing
2024-03-21 14:39:21 -03:00
fe50a1711b
fix(ds-egress): drop pending batch on failures
...
Before this commit, messages in the current batch will be retried as
part of next batch. This could have led to message duplication which is
probably not what the user wants by default.
2024-03-20 13:20:25 +01:00
a1f5de3f5b
fix(dsrepl): turn memoize into a safer function
2024-03-20 13:20:24 +01:00
d39ca41070
chore(dsrepl): mark per-node `add_generation` RPC target obsolete
...
Also annotate internal exports with comments according with their
intended use.
2024-03-20 13:20:24 +01:00
35b18f9125
fix(dsrepl): properly handle error conditions in generation mgmt
...
Also update few outdated typespecs. Also make error reasons easier
to comprehend.
2024-03-20 13:20:24 +01:00
f2268aa69a
fix(dsrepl): use correct base dir for ra system stuff
...
Co-Authored-By: Thales Macedo Garitezi <thalesmg@gmail.com>
2024-03-20 13:20:24 +01:00
404e919494
refactor(dsrepl): make shard allocator more robust and consistent
...
Co-Authored-By: Thales Macedo Garitezi <thalesmg@gmail.com>
2024-03-20 13:20:24 +01:00
0e18bd6e80
fix(dsrepl): increase replication site id bitsize back
...
In order to minimize chances of site id collision to practically zero.
2024-03-20 13:20:24 +01:00
ac9700dd28
fix(dsrepl): split shard allocator into a separate module
2024-03-20 13:20:23 +01:00
1b647035d0
chore(dsrepl): make dialyzer a bit happier
2024-03-20 13:20:23 +01:00
611b3f0e07
feat(dsrepl): use more straightforward way to drop ra shards
2024-03-20 13:20:23 +01:00
74881e8706
feat(dsrepl): make storage layer unaware of granularity of time
...
Storage also becomes a bit more pure, depending on the upper layer to
provide the timestamps, which also makes it possible to handle more
operations idempotently.
2024-03-20 13:20:23 +01:00
3cb36a5619
feat(ds-lts): extract timestamp from storage key itself
...
1. This avoids the need to deserialize the message to get the timestamp.
2. It also makes possible to decouple the storage key timestamp from the
message timestamp, which might be useful for replication purposes.
2024-03-19 20:21:56 +01:00
5cc0246351
feat(dsrepl): allow to tune select ra options
2024-03-19 20:21:55 +01:00
54b5adf868
feat(dsrepl): allocate shards predictably
...
To ensure strictly optimal and fair shard allocation across
cluster. Before this commit it was quite easy to end up with
an allocation significantly skewed towards some node, because
of the nature of randomness and relatively small number of
shards.
2024-03-19 20:21:55 +01:00
887e151be5
fix(dsrepl): handle errors gracefully in shard egress process
...
Also add cooldown on timeout / unavailability.
2024-03-19 20:21:53 +01:00
e16aee99b4
chore(dsrepl): clarify how to perform leadership transfer in runtime
2024-03-19 20:21:18 +01:00
00d509f27b
feat(dsrepl): prefer local replica in read path
...
To optimize out any unnecessary RPCs. Given the load should be
smoothed evenly across the cluster, choosing non-leader node is
not a priority.
2024-03-19 20:11:42 +01:00
19305c223c
fix(dsrepl): require majority for replication-related tables
2024-03-19 20:11:42 +01:00
f89909f60c
fix(dsrepl): tolerate trigger election timeouts for existing servers
2024-03-19 20:11:42 +01:00
3b59cf2ebf
feat(dsrepl): move shard allocation to separate process
...
That starts shard and egress processes only when shards are fully
allocated.
2024-03-19 20:11:41 +01:00
4dafbf21f6
fix(dsrepl): make db + shard part of machine state
...
It doesn't feel right, but right now is the easiest way to have it
in the scope of `apply/3`, because `init/1` doesn't actually invoked
for ra machines recovered from the existing log / snapshot.
2024-03-19 20:11:41 +01:00
d19128ed65
feat(dsrepl): cache shard metadata in persistent terms
2024-03-19 20:11:41 +01:00
e6c2c2fb07
feat(dsrepl): manage generations / db config through ra machine
2024-03-19 20:11:39 +01:00
5e94bdb932
feat(dsrepl): allocate shards once predefined number of sites online
...
Before this commit the most likely shard allocation outcome was
that all shard are allocated to just one node.
2024-03-19 20:11:03 +01:00
be793e4735
fix(dsrepl): reassign timestamp at the time of submission
...
This is needed to ensure total message order for a shard, and
guarantee that no messages will be written "in the past". which
may break replay consistency.
2024-03-19 20:11:01 +01:00
146f082fdc
feat(dsrepl): implement raft-based replication
...
Still very rough but mostly working.
2024-03-19 20:09:44 +01:00
f2dc940436
Merge remote-tracking branch 'upstream/release-56' into 0319-sync-release56
2024-03-19 15:20:08 +01:00
11ae04d810
test(ds): rm unused var warning
2024-03-14 17:16:24 -03:00
2ebc8dcc55
fix(ds): use `infinity` timeout when storing batches
2024-03-14 10:17:18 -03:00
6af01b916e
feat(ds): implement `get_delete_streams`, `make_delete_iterator` and `delete_next` callbacks for builtin storage
...
Part of https://emqx.atlassian.net/browse/EMQX-11841
2024-03-08 09:56:46 -03:00
f7e3afde16
test(ds): avoid introducing new macros
2024-03-07 16:49:20 +01:00
69427dc42d
test(ds): add tests for error mapping and replay recovery
2024-03-07 12:59:58 +01:00
09905d78cd
chore(ds): make error handling slightly simpler
...
Co-Authored-By: Thales Macedo Garitezi <thalesmg@gmail.com>
2024-03-07 12:59:57 +01:00
b39c710ec2
fix(ds): tidy up few typespecs
2024-03-07 12:59:57 +01:00
2146d9e1fe
feat(ds): introduce error classes in critical API functions
...
For now, only recoverable / unrecoverable errors are introduced.
2024-03-07 12:59:57 +01:00
5d87d400f4
feat(ds): add atomic store API
...
Part of https://emqx.atlassian.net/browse/EMQX-11841
2024-03-06 15:24:14 -03:00
06334798a5
fix(ds): fix `drop_generation` typespec
...
This typespec fix will be used downstream by other backends.
2024-03-04 14:15:59 -03:00
b706caf294
feat(ds): export types
2024-02-29 14:27:18 +03:00
d5ae0e5c53
feat(ds): update delete/count interface
2024-02-28 22:51:24 +03:00
b010d34640
chore(ds): add delete callbacks
2024-02-26 17:35:13 +03:00
46877e979b
chore: update copyright-year
2024-02-23 08:21:06 +01:00
d469f4158e
chore: bump app vsns
2024-02-20 16:53:57 -03:00
8cfb22f0b8
fix(ds): Retry getting the shard leader
2024-02-16 12:42:48 +01:00
280fcd8c52
Merge pull request #12437 from ieQu1/dev/optimize_make_filter
...
Optimize emqx_ds_bitmask_keymapper:make_filter function.
2024-02-05 17:32:28 +01:00
4665837cf0
fix(ds): Apply review remarks
2024-02-05 16:52:06 +01:00
c7888ad1f1
Merge pull request #12475 from ieQu1/dev/lean-stream
...
Use a more compact data structure to represent streams
2024-02-05 13:55:24 +01:00
698ba3f271
fix(ds): Optimize emqx_ds_bitmask_keymapper:make_filter
...
This optimization makes idle polling faster
2024-02-05 10:54:19 +01:00
Thales Macedo Garitezi