ThinkUp’s database grows very quickly, especially if an installation has several active service users set up in it. If your database is so big and unwieldy that your server slows to a crawl while running migrations–and maybe doesn’t ever complete them–it’s time to reduce the size of your database.
The easiest way to to reduce the size of a big ThinkUp installation’s database is to separate out data-heavy service users to other databases/installations. To do so, use the service user export tool to download each individual user’s archive.
Then, set up new ThinkUp installations and import that data into them.
The first issue was that we had the key_buffer_size and myisam_sort_buffer_size set way too low. We had them both set at 8MB, I believe (probably just MySQL defaults). I upped those values to 1GB each (on a machine with 4GB RAM). MySQL isn’t guaranteed to use the full 1GB of RAM for each, though.
The key_buffer_size should be somewhere between 25% and 50% of the total RAM of the machine (assuming that is reasonable based on other services on the machine):
I believe that the myisam_sort_buffer_size had a much bigger impact on the performance of the alter tables, though:
Specifically, this section:
“If you use any option to ALTER TABLE other than RENAME, MySQL always creates a temporary table, even if the data wouldn’t strictly need to be copied (such as when you change the name of a column). For MyISAM tables, you can speed up index re-creation (the slowest part of the alteration process) by setting the myisam_sort_buffer_size system variable to a high value.”
Once I had those in place, I could watch the temp table as the index built much faster. Once it got past creating the temp table, however, it went into “Repair with keycache”, which is un-fast:
Ideally, you want MySQL to use “Repair with sorting” (which uses the MySQL tmpdir) when doing an index rebuild for an alter. In order to use “Repair with sorting”, there must be enough room in the tmpdir to store a small factor (2-5?) times the size of the total indexes. In our case, we had the tmpdir set to /tmp, which is only a 2GB partition. 2GB wasn’t enough space for the index, so it was falling back to “Repair with keycache”. I switched our tmpdir to a different directory/partition with more space and tried again. I was still running into “Repair with keycache”. This time, the issue was the myisam_max_sort_file_size. It defaults to 2GB, but 2GB wasn’t enough for our follows table. Since we have sufficient disk, I upped this value up to 32GB. Then, MySQL was able to properly use “Repair with sorting”. At that point, the alter table took 45 minutes vs. the unknown amount of time it was going to take prior to making any of these changes (10+ hours just to build half of the temp table before I killed it).