WAL Bypass for indexes

I have followed the discussion from 3 months ago on WAL bypass and wanted to offer some more information.

I have long believed that the bottleneck in transaction-oriented systems is the writing of the indexes, complete with splits and merges. A single update to one field of a heavily-indexed table could cause dozens of index writes to cascade.

The point is that speeding up index writing should offer the most "bang for the buck" relative to performance. As a bonus, a corrupted index can ALWAYS be recovered, while a corrupted table cannot.

I did some informal testing using pgbench on v8.07. First, I ran pgbench normally with 75 users doing 100 transactions, full vacuuming between runs. My machine consistently gave me 92 tps.

As an experiment, I commented out of the btree index source all of the XLOG code I could find. I basically replaced the test for a temp table with "if (0)" and then recompiled.

Running again pgbench with 75 users and 100 transactions, I received a consistent rate of 132 tps, a 40% increase in throughput.

It seems to me that major performance gains can be had by allowing some indexes to be created with some "UNSAFE-FAIL" flag, which would eliminate WAL logging and fsync() for the index files.

Upon recovery, the index gets rebuilt. The only downside is potentially long rebuild times during recovery.

Thoughts?

Sincerely,
Marty