[POC - do not review yet] chore: almost mutex free replication #5758
Conversation
src/server/server_family.cc
Outdated
| }; | ||
| fb2::LockGuard lk(replicaof_mu_); | ||
| // Deep copy because tl_replica might be overwritten inbetween | ||
| auto replica = tl_replica; |
There was a problem hiding this comment.
Bread and butter no1 of this PR. Bye bye blocking info command because of the mutex.
There was a problem hiding this comment.
TODO worth adding a test
| // If we are called by "Replicate", tx will be null but we do not need | ||
| // to flush anything. | ||
|
|
||
| util::fb2::LockGuard lk(replicaof_mu_); |
There was a problem hiding this comment.
Bread and butter no2. No more loading state prematurely. We only get into loading_state if we are doing full sync. Otherwise, no state change at all.
|
TODO: move this to |
kostasrim
force-pushed
the
kpr3
branch
3 times, most recently
from
38640b3 to
76598f2
Compare
Signed-off-by: kostas <[email protected]>
|
|
||
| logging.info(f"succeses: {num_successes}") | ||
| assert COMMANDS_TO_ISSUE > num_successes, "At least one REPLICAOF must be cancelled" | ||
| assert COMMANDS_TO_ISSUE == num_successes |
There was a problem hiding this comment.
The new algorithm does not use two phase locking so the following is no longer a possibility:
- client 1 -> REPLICAOF -> Locks the mutex -> updates
replica_tonew_replica-> releases the mutex -> calls replica_->Start() - client 2 -> REPLICAOF -> same as (1) but first calls replica_->Stop() -> releases the mutex
- client 1 -> REPLICAOF command grabs the second lock to the mutex, observes that the context got cancelled because of step (2) and boom returns "replication cancelled"
This can not happen anymore because we lock only once and atomically update everything including stopping the previous replica. So by the time (2) grabs the lock in the example above, the previous REPLICAOF command is not in some intermediate state. To observe that indeed we cancelled, we should read the logs and see ("Stopping replication") COMMANDS_TO_ISSUE - 1 times + 1 (because of the Shutdown() at the end)
Bonus points is that I suspect we might be able to also solve #4685 but I will need to follow up with that
Signed-off-by: kostas <[email protected]>
Do not review.