The bug was introduced when I changed the behaviour of conf_change in
regard to instances with target_state Expelled. As a result we would
sometimes arbitrarily demote healthy voters in presence of degenerate
ones.

For example we could have this situation:

instance i1: raft_id=1, target_state=Online, raft_configuration=voter
instance i2: raft_id=2, target_state=Expelled, raft_configuration=voter (!)
instance i3: raft_id=3, target_state=Offline, raft_configuration=voter (!)
instance i4: raft_id=4, target_state=Online, raft_configuration=learner (!)
instance i5: raft_id=5, target_state=Online, raft_configuration=learner (!)

Dmitry Rodionov authored 2 months ago and Егор Ивков committed 2 months ago

The check arrived late when elections were already completed.
Circumstances are a bit weird. For some reason the check hang for two
seconds. Only after that it received conn refused, and during these two
seconds elections happened and leader was chosen so check can never
succeed after that.

Artur Sabirov authored 2 months ago and Maksim Kaitmazian committed 2 months ago

Follow-up for !1425

BREAKING CHANGE: as soon as PostgreSQL protocol
is enabled by default, we must be careful with
`--pg-listen` cli flag when creating more than
a single instance, because we will suddenly get
a port confict error (busy port)

Logs help to debug flaky tests. Because sometimes just by instance logs
it is not that easy to guess what the test was doing at that time.
Especially when there is a repetitive pattern in the test.

For now instance logs and test output are printed in separate sections.
It still allows to match events by time easier compared to looking only
at instance log (it is way bigger compared to test log).

Additionally this patch sets up a `framework` package intended to host
reusable test code without polluting conftest.py

I found this by compiling tarantool with CMAKE_BUILD_TYPE=Debug

We used to automatically truncate the index in compact_log if the caller
requested to compact too many entries. This made it so that the
requirement of not compacting any un-applied entries was implicit in our
code base, which is not good as it allows for some bugs to creep in
(like the one we fix a couple commits ago).

Now this is changed and instead of silently adjusting the index of last
compacted entry, we just assert that it's no greater than the applied
index. As a consequence there's a minor improvement in do_raft_log_auto_compaction
function.

There was a hard-to-reproduce bug in our snapshot application code.
We always compact the raft log before applying the snapshot, because
the snapshot replaces the entries and some of the logic in raft-rs
seems to rely on this. The problem was, that our compact_log function
would not remove any unapplied entries, which makes sense for compaction
triggered automatically by raft log size, but doesn't make sense for
raft snapshot, as the snapshot contains the state corresponding to the
newer entries. The fix is simple: don't guard from unapplied entry
compaction in case the compaction is for raft snapshot.

We don't add any regression tests for this, because the implementation
would be too difficult and would need us to pollute the code with error
injection logic, which is not a worthy trade off in this case. But also
the logic will still be tested, because this bug was responsible for a
large amount of flaky tests, so we should see a significant reduction in
flakiness from now on in tests concerning raft snapshots.

Maksim Kaitmazian authored 2 months ago and Maksim Kaitmazian committed 2 months ago

Close #1228

TopologyCache is a collection of deserialized structures with
information about cluster topology. This currently includes data from
_pico_instance, _pico_replicaset, _pico_tier & _pico_service_route_table.
The info is automatically kept up to date with the corresponding system
tables. The TopologyCache also caches the immutable info related to the
current instance, like instance name, replicaset uuid, etc.

From now on we should be reading this data from TopologyCache whenever
possible instead of going directly to system tables as we were doing
previosly.

At the moment only the plugin RPC module has transitioned to using
TopologyCache but other modules should be refactored as well.
Especially governor.

Before this fix if requesting RPC by an invalid bucket_id we would send
an RPC to every replicaset to check if they have such a bucket_id.
This is not needed because we know the allowed range.

this adr is a proposal of implementing TRUNCATE sql command, its context and possible problems