Sunday, May 15, 2016

rom Indexes and Search

So... some updates on rom.

The end of January, I posted a teaser picture about "Interleaved Indexes" in rom/Redis via tweet. If you haven't seen the picture, it's here:


Interleaved index?

I ended up building an interleaved index in Redis using a bit of Lua scripting and Python. No ZRANGEBYLEX, surprisingly. What is an interleaved index? It's what I was calling an indexing mode that has the same ordering and filtering options as a few different multi-dimensional indexes stored as tree-structures. Specifically some KD trees, Redshift interleaved sort keys, BSP trees, a specific implementation of crit-bit trees, and several others offer the specific functionality I was after.

Why

The simple reason why I was implementing an interleaved index was because I see some intersection options on data in Redis to be potentially less efficient than a proper multi-dimensional index would or could be. Long story short, it worked, but not without issues. I mentioned some of these issues in a series of tweets 1, 2, and 3, semi-abandoned the code in late-February, and now am ultimately not releasing it. Why? Because it doesn't actually add anything. It was complex, borrowed about 750 lines of code I wrote 5 1/2 years ago, and ... no.

A better option

There were a few very simple wins that I knew could be made with the query optimizer, including a fix on my side for a bug when calling TYPE from within a Lua script (which returns a table instead of a string). The ultimate result of that work is that some queries on numeric ranges can be hundreds or thousands of times faster on large indexes in theory. Partly due to starting with the correct set/sorted set to start, but also implementing a direct scan of an index instead of intersect/union + delete outside ranges.

Sometimes being more specific for optimizations is worth it. Definitely is in this case. For one of my use-cases involving search, I'm seeing 10-20x faster queries in practice, and 150x faster in a simple canned test.

I also removed non-Lua writing mode code. Sorry for those of you living in pre-2.6 days, but you'll have to upgrade. Hell, even with Lua scripting turned off, the query optimizer still used Lua, so if this worked in Redis 2.4 recently, I'd be surprised.

So that's what's going on right now.

Rebuild your indexes

Yeah. And rebuild your indexes. I'm sorry. Whenever I'm using rom as a cache or index of some kind, I re-cache and re-index daily so things like this always eventually resolve themselves, especially immediately after a library upgrade. Not a new idea; Google did it with their bigtables for cleanup, Postgres does auto-vacuum. Call this a manual vacuum via re-indexing.

Once/day or however often, maybe during off hours:

# import all of your models first
# then...
from rom import columns, util
for model in columns.MODELS.values():
    util.show_progress(util.refresh_indices(model))

That will rebuild all of the indexes on all of your models.

Almost P.S. - Loadable modules in Redis?

Redisconf 2016 happened last week and loadable modules were announced. I think that for people who host their own Redis, it could be awesome. Think of it like an answer to Postgres plugins. Hope I can pay someone to run my loadable modules, if I ever get around to building any :P

1 comment: