Non-technical summary: things are lots faster at Tizra sites and admin tools. There's certainly more to do, but we've got more tricks up our sleeves! Because the big current speed boost is related to one cause, and it took me a while to track down, the geek appendage to this post describes what we found and how we fixed it.
I spent a bunch of time last week looking at system performance. As we've been adding customers and usage, we were beginning to feel the pinch. Performance always varies, but the range of response times was getting wider as things slowed, leading me to think that there might be some systemic issues that would give us a quick improvement (and indeed there was some Linux tuning that helped a bit). But data access seemed to be the real issue, so I spent a bunch of time looking into hibernate, and our caching and querying, and then wound up spending a day or so basically watching all the queries go through Postgres. And you know what? most of them seemed much slower than they should be, even though they are pretty hairy.
Of course, the next step was to check for database indexes, and how the query plans were using them. But in hand testing the plans looked good, and the indexes were sensible. But when run by hand the queries were also significantly faster than when hibernate ran them! This was much easier to see now that we have a live load, which is inevitably different from a test setup. So why the difference? Postgres was ignoring our indexes only when Tizra publisher made the queries.
Turns out that there's an old bug in Postgres where it would ignore indexes on bigint fields in prepared statements unless there was an explicit data type cast. (That type confusion was an obscure result of skew between Postgresql and the SQL standard.) And that was the behavior I was seeing, even though we were using a much more recent vintage of all the software. This was terrible for us, because we have a multi-tenant publishing system for large document collections and we use bigints as primary object identifiers!
So, why the old problem if the bug is gone, and we are not using postgres 7? It turns out that we dynamically build those hairy queries, in HQL (hibernate query language), using the String trick. But nowadays instead of making your indexes work, it breaks them! The differences are invisible in the SQL. It turned out that we were in a version "donut hole." Our database was recent enough so the String trick worked the opposite way (preventing fast queries for our prepared statements), but the JDBC driver wasn't making the calls in the right way to make the old trick work. End result: we're now running the latest JDBC driver with compatibility options set while we update our hairy query generator. And now we can really start tuning our setup!
If the web had not provided the history of the old bug, I would have had a much worse time even knowing where to look to find our somewhat subtle configuration issue. So enjoy the speedup, I sure am!