Test: fault tolerance at master side

ttyusupov commented 3 years ago

Jira Link: DB-2359 From https://phabricator.dev.yugabyte.com/D8253#243515: " We should also consider some unhappy paths:

crash or leader failover before sending out the split tasks
long enough partition while trying to send out the splits so that they timeout and not get executed

We'll have some tablets in CREATING state, which have their table_id set, but the split does not get retried. For a timeout, they're also not added to the TableInfo object at all. For the failover case, I believe they'll automatically be added to the tablet, using AddTablet (in the tablet loader code). "

bmatican commented 2 years ago

@SrivastavaAnubhav can you chime in if we have coverage by now for all of these?

SrivastavaAnubhav commented 2 years ago

These should be handled by automatic tablet splitting, if it's enabled (but it's not enabled by default yet).

bmatican commented 2 years ago

@SrivastavaAnubhav oh, are you saying that if automatic is turned off, but a user does manual splits, we might not be re-triggering splits after master failover, if neeed? If so, might be worth discussing in next week's sync, as that seems a bit off.

SrivastavaAnubhav commented 2 years ago

Yeah, we should have some common method that gets called on restart.

yugabyte / yugabyte-db

Test: fault tolerance at master side #8645