Got my app up and running under 3.1 well enough to start doing some benchmarking/profiling runs. One hot spot that immediately jumped out in the profiler was JdbcTokenStore.insertOrUpdateToken(). It seems that updating a ResultSet is quite slow in the PostgreSQL JDBC driver; under the hood, for each field modification, the JDBC driver ends up querying the table metadata from the database to find out the table's primary key so it can construct a plain UPDATE statement which it then executes. It doesn't appear to cache the table metadata across queries, plus the three resultSet.update*() calls in that method end up causing three separate metadata lookups and three separate UPDATEs.
Subclassing JdbcTokenStore to override that method and replace it with a plain "UPDATE TokenStore SET token = ?, tokenType = ?, timestamp = ? WHERE processorName = ? AND segment = ?" gave me an immediate 13% throughput gain on my app as a whole in my test environment. I got another 5% gain from doing the same thing to claimToken().
It's quite likely this is database-specific and that on other databases, it's faster to update a ResultSet. And the benefit will vary depending on the number of tracking processors in the app. But for Postgres, at least, there's an easy win here.