It is the exciting return of the global checkpoint background
sync. Long, long ago, in snapshot version far, far away we had and only
had a global checkpoint background sync. This sync would fire
periodically and send the global checkpoint from the primary shard to
the replicas so that they could update their local knowledge of the
global checkpoint. Later in time, as we sped ahead towards finalizing
the initial version of sequence IDs, we realized that we need the global
checkpoint updates to be inline. This means that on a replication
operation, the primary shard would piggy back the global checkpoint with
the replication operation to the replicas. The replicas would update
their local knowledge of the global checkpoint and reply with their
local checkpoint. However, this could allow the global checkpoint on the
primary to advance again and the replicas would fall behind in their
local knowledge of the global checkpoint. If another replication
operation never fired, then the replicas would be permanently behind. To
account for this, we added one more sync that would fire when the
primary shard fell idle. However, this has problems:
- the shard idle timer defaults to five minutes, a long time to wait
for the replicas to learn of the new global checkpoint
- if a replica missed the sync, there was no follow-up sync to catch
them up
- there is an inherent race condition where the primary shard could
fall idle mid-operation (after having sent the replication request to
the replicas); in this case, there would never be a background sync
after the operation completes
- tying the global checkpoint sync to the idle timer was never natural
To fix this, we add two additional changes for the global checkpoint to
be synced to the replicas. The first is that we add a post-operation
sync that only fires if there are no operations in flight and there is a
lagging replica. This gives us a chance to sync the global checkpoint to
the replicas immediately after an operation so that they are always kept
up to date. The second is that we add back a global checkpoint
background sync that fires on a timer. This timer fires every thirty
seconds, and is not configurable (for simplicity). This background sync
is smarter than what we had previously in the sense that it only sends a
sync if the global checkpoint on at least one replica is lagging that of
the primary. When the timer fires, we can compare the global checkpoint
on the primary to its knowledge of the global checkpoint on the replicas
and only send a sync if there is a shard behind.
Relates #26591
f35d1de502