postgres_driver: better sync lock in partition creation (#2809)

With the original approach it happened cases where one connection
acquired the lock and other connection tried to release the same lock,
causing an unrecoverable failure which made the node to end.
This commit is contained in:
Ivan FB 2024-06-14 10:07:41 +02:00 committed by GitHub
parent 0561e5bd45
commit ab4b5b27bd
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 61 additions and 16 deletions

View File

@ -70,6 +70,9 @@ proc addPartitionInfo*(
trace "Adding partition info"
self.partitions.addLast(partitionInfo)
proc clearPartitionInfo*(self: PartitionManager) =
self.partitions.clear()
proc removeOldestPartitionName*(self: PartitionManager) =
## Simply removed the partition from the tracked/known partitions queue.
## Just remove it and ignore it.

View File

@ -866,6 +866,11 @@ proc acquireDatabaseLock*(
s: PostgresDriver, lockId: int = 841886
): Future[ArchiveDriverResult[void]] {.async.} =
## Acquire an advisory lock (useful to avoid more than one application running migrations at the same time)
## This should only be used in the migrations module because this approach doesn't ensure
## that the lock is acquired/released by the same connection. The preferable "lock"
## approach is using the "performWriteQueryWithLock" proc. However, we can't use
## "performWriteQueryWithLock" in the migrations process because we can't nest two PL/SQL
## scripts.
let locked = (
await s.getStr(
fmt"""
@ -908,6 +913,46 @@ proc performWriteQuery*(
return ok()
const COULD_NOT_ACQUIRE_ADVISORY_LOCK* = "could not acquire advisory lock"
proc performWriteQueryWithLock*(
self: PostgresDriver, queryToProtect: string
): Future[ArchiveDriverResult[void]] {.async.} =
## This wraps the original query in a script so that we make sure a pg_advisory lock protects it
debug "performWriteQueryWithLock", queryToProtect
let query =
fmt"""
DO $$
DECLARE
lock_acquired boolean;
BEGIN
-- Try to acquire the advisory lock
lock_acquired := pg_try_advisory_lock(123456789);
IF NOT lock_acquired THEN
RAISE EXCEPTION '{COULD_NOT_ACQUIRE_ADVISORY_LOCK}';
END IF;
-- Perform the query
BEGIN
{queryToProtect}
EXCEPTION WHEN OTHERS THEN
-- Ensure the lock is released if an error occurs
PERFORM pg_advisory_unlock(123456789);
RAISE;
END;
-- Release the advisory lock after the query completes successfully
PERFORM pg_advisory_unlock(123456789);
END $$;
"""
(await self.performWriteQuery(query)).isOkOr:
debug "protected query ended with error", error = $error
return err("protected query ended with error:" & $error)
debug "protected query ended correctly"
return ok()
proc addPartition(
self: PostgresDriver, startTime: Timestamp
): Future[ArchiveDriverResult[void]] {.async.} =
@ -930,21 +975,15 @@ proc addPartition(
"CREATE TABLE IF NOT EXISTS " & partitionName & " PARTITION OF " &
"messages FOR VALUES FROM ('" & fromInNanoSec & "') TO ('" & untilInNanoSec & "');"
# Lock the db
(await self.acquireDatabaseLock()).isOkOr:
error "failed to acquire lock", error = error
return err("failed to lock the db")
defer:
(await self.releaseDatabaseLock()).isOkOr:
error "failed to release lock", error = error
return err("failed to unlock the db.")
(await self.performWriteQuery(createPartitionQuery)).isOkOr:
(await self.performWriteQueryWithLock(createPartitionQuery)).isOkOr:
if error.contains("already exists"):
debug "skip create new partition as it already exists: ", skipped_error = $error
return ok()
if error.contains(COULD_NOT_ACQUIRE_ADVISORY_LOCK):
debug "skip create new partition because the advisory lock is acquired by other"
return ok()
## for any different error, just consider it
return err(fmt"error adding partition [{partitionName}]: " & $error)
@ -953,9 +992,12 @@ proc addPartition(
self.partitionMngr.addPartitionInfo(partitionName, beginning, `end`)
return ok()
proc initializePartitionsInfo(
proc refreshPartitionsInfo(
self: PostgresDriver
): Future[ArchiveDriverResult[void]] {.async.} =
debug "refreshPartitionsInfo"
self.partitionMngr.clearPartitionInfo()
let partitionNamesRes = await self.getPartitionsList()
if not partitionNamesRes.isOk():
return err("Could not retrieve partitions list: " & $partitionNamesRes.error)
@ -994,13 +1036,13 @@ proc loopPartitionFactory(
debug "starting loopPartitionFactory"
## First of all, let's make the 'partition_manager' aware of the current partitions
(await self.initializePartitionsInfo()).isOkOr:
onFatalError("issue in loopPartitionFactory: " & $error)
while true:
trace "Check if we need to create a new partition"
## Let's make the 'partition_manager' aware of the current partitions
(await self.refreshPartitionsInfo()).isOkOr:
onFatalError("issue in loopPartitionFactory: " & $error)
let now = times.now().toTime().toUnix()
if self.partitionMngr.isEmpty():