Affects Version/s: None
Fix Version/s: None
HBASE-7051 lists 3 options for fixing an ACID-violation wrt checkAndPut:
1) Waiting for the MVCC to advance for read/updates: the downside is that you have to wait for updates on other rows.
2) Have an MVCC per-row (table configuration): this avoids the unnecessary contention of 1)
3) Transform the read/updates to write-only with rollup on read.. E.g. an increment would just have the number of values to increment.
Another option occurred to me that I think is worth investigating: rely on a row-level read/write lock rather than MVCC.
Here is pseudo-code for what exists today for read/updates like checkAndPut
Write-only operations (e.g. puts) are the same, just without step 1a.
Now, consider the following instead:
As before, write-only operations are the same, just without step 1a.
The difference here is that writes grab a row-level read lock and hold it until the MVCC is completed. The nice property that this gives you is that read/updates can tell when the MVCC is done on a per-row basis, because they can just try to acquire the write-lock which will block until the MVCC is competed for that row in step 7.
There is overhead for acquiring the read lock that I need to measure, but it should be small, since there will never be any blocking on acquiring the row-level read lock. This is because the read lock can only block if someone else holds the write lock, but both the write and read lock are only acquired under the row lock.
I ran a quick test of this approach over a region (this directly interacts with HRegion, so no client effects):
- 30 threads
- 5000 increments per thread
- 30 columns per increment
- Each increment uniformly distributed over 500,000 rows
- 5 trials
Better-Than-Theoretical-Max: (No locking or MVCC on step 1a): 10362.2 ms
Today: 13950 ms
The locking approach: 10877 ms
So it looks like an improvement, at least wrt increment. As mentioned, I need to measure the overhead of acquiring the read lock for puts.