Vladislav Shpilevoy
Vladislav Shpilevoy
### Bug 5 - WAL queue crash In this example I schedule the queued WAL fibers in the wrong order. ```Lua -- -- Step 1 -- fiber = require('fiber') log...
### Bug 6 - Outdated read-view being sent This is a tricky one. And is not dependent on the bugs above. I.e., fixing them all still keeps this new bug...
### Bug 7 - Join doesn't follow the transaction execution order and can miss rollbacks Here is a simple diff which is applied on top of the current fix branch:...
### Bug 8 - Duplicate entries being sent In `memtx_engine_join()` the call `wal_sync()` gets vclock of the last txn which is in the read-view. It assumes that all the txns...
@sergepetrenko for that bug, if the problem is about `wal_sync()`, we would need its out-vclock to be higher than what is included into the read-view. I am not sure I...
### Bug 9 - Duplicate `CONFIRM` Not all bugs in this ticket are reproducible on master (many are time bombs in the code, waiting to get exposed when other bugs...
@drewdzzz I have no new comments, good work! Also thanks to @Serpentian for his notes. Now we just need to finalize those small things.
The crash is not related to DDL. Apparently, it happens on shutdown when the limbo still has a txn. And that is possible, if shutdown is started when a transaction...
The crash is fixed. Now we only work on the replication blockade.
To be clear - this is only for recovery speedup, right? At runtime when I create a new secondary index, I still will iterate over the primary index and insert...