Re: large tables, large keycaches

[Mike Connell]

Hi,

What you are saying about the disk being the bottleneck is exactly the problem with this
situation. The key buffer is acting like an in-memory cache until it fills up. After that, its
how fast can the disk go.

I'm seeing i/o writes of 4Kbytes, and i/o reads of 4Kbytes. I was hoping to get our Ops guys
to change the stripe size to this value. Other than that, would key_cache_block_size be of
any use? It is currently set at 1024 which doesn't match what the 4K value the OS is actually
writing to disk (using iosnoop output as the monitoring tool).

Your 2nd statement is exactly what I was hoping to simulated by writing a CRON job
to do a 'flush table mybigtable' periodically.

This won't work currently since its all or nothing with flush table on really big tables.
In the future, a special variant of flush table that could do *block* flushing of least recently
used key buffers would be very valuable in this case IMHO. A seperate bg thread would
be even better.

Thanks