...
Hold lock on serverlist, compute mappings, add mappings to local coordinator state, release lock, log to logcabin. Note about timestamp (headoflog): We can safely put the headoflog as 0 while creating a table. This means that during table creation, i don't need to talk to masters to get the headoflog at all.
Better solution 5: Solution 2 just works with this realization!
During the entire create table operation (or most coordinator write operations for that matter), i have a table manager lock. This means that while i am doing the operation, if a master had crashed, when its recovery wants to read the tablet map from table manager, it will have to stall till this table creation has completed. If the coordinator had failed and is recovering the create table operation, then i anyway can't do master recovery in parallel since i'm first going to complete coordinator recovery before servicing rpcs. So solution 2 just works with this realization (the idea for which came from solution 4).
This works!
| Anchor | ||||
|---|---|---|---|---|
|
...