A multi-column index on (a, b, c) can't be used for a query that filters on b and c but not a. More detailed documentation here: http://www.postgresql.org/docs/9.3/static/indexes-multicolum...
Adding N 1-column indexes for N columns on every table is at best is a lazy cop-out for actual data and query analysis.
Edit: to address your example specifically, you'd probably be better off having two indexes, one on (a, b, c) and another (b, c) instead of three separate indexes if that covers your query patterns.
Let's say you care about ten different types of "high value" events, which reference a total of six columns. I'll assume we can cover them with three multi-column indexes, although it wouldn't be hard to cook up a realistic scenario in which you'd need more. That means, for each INSERT/UPDATE, you need to write to three different indexes.
Given that the event definitions are selective, a single partial index requires a write for about 0.05% of INSERT/UPDATEs. Ten of these will cumulatively require one write on ~0.5% of inserts -- a 600x improvement over the conservative estimate above. That is, the cost of of maintaining the set of multi-column indexes should be much, much higher than the cost of maintaining the set of partial indexes.
As I mentioned in the article, the partial index approach also allows a more flexible set of predicates. What if you want to index for rows with a field that matches a fixed regex?
Edit: Apparently I can't respond to your response -- does HN have a chain length limit?
Yes -- for indexing the single event definition used for the profiling here, a multi-column index would absolutely be preferable to three single-column indexes. I didn't think to include the option because I assumed it wouldn't scale to a case in which you have ten different event types that use an overlapping but not identical set of fields. Definitely would have been a good idea to include a comment to this effect in the post.
And if you reread my first comment, I fully acknowledge that the write performance of partial indexes would likely still be better than a single full index.
That makes a bit more sense. I would challenge your phrases "realistic scenarios," "assumed it wouldn't scale," and "overlapping but not identical set of fields," but the real problem we're facing in this thread is all our scenarios are hypothetical and simplified. The fact is, indexing decisions are highly dependent on the exact scenario and since we have different contexts the best we could do is banter on about the "most accurate" hypothetical scenario, which is nonsense.
So I guess my conclusions would be "highly selective partial indexes could save you a bunch of write io" and "otherwise, It Depends™ because Indexes Are Hard®". In any case, your article is a decent introduction to partial indexes in general.
Agreed in general. No one-size-fits-all answers, and schematizing your data well is going to require case-by-case attention and experimentation for the foreseeable future.
The multi-column index takes up 755 mb, vs 1026 mb for the set of three single-column indexes.
If you're always going to query those fields in that order, that's a perfect way to do things. If you'll be querying it a couple different ways then you're way better off with three separate one column indexes assuming the RDBMS can make use of index intersection.