When remote DC is down, the remote team collection of DD can initializing
waiting for the remote to recover (all_tlog_recruited state). However, the
getTeam request can already be served by the remote team collection. So, for
a RelocateShard (data movement such as split, move), it will get a team for
the remote DC. But the data movement can't make progress on the remote team
because the remote DC hasn't recovered yet. Because of the stuck of data
movement, the primary cannot reach the "storage_recovered" state and stay in
accepting_commit state.
The specifc test failure: slow/ApiCorrectness.toml -s 339026305 -b on
at commit: 0edd899d65
In this test, primary DC has 1 SS killed, remote DC has 2 TLog and 2 SS killed.
So the remote is dead, the remaining 2 SSes can't make progress because of the
loss of 2 TLogs. The repairDeadDatacenter() can't reach the "storage_recovered"
state due to DD's failure of moving shards away from the killed SS in the
primary.
The fix is to exclude all remote in repairDeadDatacenter() so that tells DD to
mark all SSes in the remote as unhealthy. Another fix is to return empty
results for getTeam request if the remote team collection is not ready. This
will allow the data movement to continue, essentially remote team is not changed
for the data movement.
* throttle the cluster when blob workers fall behind
* do not throttle on blob workers if they are not enabled
* remove an unnecessary actor
* fixed a compile error
* fetch blob worker metrics at the same interval as the rate is updated, avoid fetching the complete blob worker list too frequently
* fixed another compilation bug
* added a 5 second delay before bw throttling to prevent false positives caused by the 100e6 version jump during recovery. Lower the throttling thresholds to react much quicker to bw lag.
* fixed a number of problems
* changed the minBlobVersionRequest to look at storage server versions since this will be a lot more efficient
* fix: do not let desired go backwards
* fix: track the version of notAtLatest changefeeds for throttling
* ratekeeper now throttled blob workers by estimating the transaction per second throughput of the blob workers
* added metrics for blob worker change feeds
* added a knob to disable bw throttling
* fixed the transaction options in blob manager
* Shard based move.
* Clean up.
* Clear results on retry in getInitialDataDistribution.
* Remove assertion on SHARD_ENCODE_LOCATION_METADATA for compatibility.
* Resolved comments.
Co-authored-by: He Liu <heliu@apple.com>
QuietDatabase will now fail by itself after 1000 seconds
instead of relying on the general simulation timeout.
Additionally it will print a more human friendly error.
* Initialize cluster version at wall-clock time
Previously, new clusters would begin at version 0. After this change,
clusters will initialize at a version matching wall-clock time. Instead
of using the Unix epoch (or Windows epoch), FDB clusters will use a new
epoch, defaulting to January 1, 2010, 01:00:00+00:00. In the future,
this base epoch will be modifiable through fdbcli, allowing
administrators to advance the cluster version.
Basing the version off of time allows different FDB clusters to share
data without running into version issues.
* Send version epoch to master
* Cleanup
* Update fdbserver/storageserver.actor.cpp
Co-authored-by: A.J. Beamon <aj.beamon@snowflake.com>
* Jump directly to expected version if possible
* Fix initial version issue on storage servers
* Add random recovery offset to start version in simulation
* Type fixes
* Disable reference time by default
Enable on a cluster using the fdbcli command `versionepoch add 0`.
* Use correct recoveryTransactionVersion when recovering
* Allow version epoch to be adjusted forwards (to decrease the version)
* Set version epoch in simulation
* Add quiet database check to ensure small version offset
* Fix initial version issue on storage servers
* Disable reference time by default
Enable on a cluster using the fdbcli command `versionepoch add 0`.
* Add fdbcli command to read/write version epoch
* Cause recovery when version epoch is set
* Handle optional version epoch key
* Add ability to clear the version epoch
This causes version advancement to revert to the old methodology whereas
versions attempt to advance by about a million versions per second,
instead of trying to match the clock.
* Update transaction access
* Modify version epoch to use microseconds instead of seconds
* Modify fdbcli version target API
Move commands from `versionepoch` to `targetversion` top level command.
* Add fdbcli tests for
* Temporarily disable targetversion cli tests
* Fix version epoch fetch issue
* Fix Arena issue
* Reduce max version jump in simulation to 1,000,000
* Rework fdbcli API
It now requires two commands to fully switch a cluster to using the
version epoch. First, enable the version epoch with `versionepoch
enable` or `versionepoch set <versionepoch>`. At this point, versions
will be given out at a faster or slower rate in an attempt to reach the
expected version. Then, run `versionepoch commit` to perform a one time
jump to the expected version. This is essentially irreversible.
* Temporarily disable old targetversion tests
* Cleanup
* Move version epoch buggify to sequencer
This will cause some issues with the QuietDatabase check for the version
offset - namely, it won't do anything, since the version epoch is not
being written to the txnStateStore in simulation. This will get fixed in
the future.
Co-authored-by: A.J. Beamon <aj.beamon@snowflake.com>
Instead, in order to enforce the maximum fault tolerance for snapshots,
update getStorageWorkers to return the number of unavailable storage
servers (instead of throwing an error when unavailable storage servers
exist).
In this PR, the blob manager now recruits blob workers
(via communication with the cluster controller). Blob workers
are onboarded as blob worker processes enter the cluster.
Evan Tschannen