Redshift is forked from ParAccel, which was originally forked from Postgres several years ago. ParAccel has had code generation (via transpilation to C) since long before Amazon even purchased its source code. In other words, Redshift has had this feature since its inception, albeit via C-transpilation (compiled with GCC) not LLVM codegen.
In addition, Redshift's primary index is a column store which is particularly well-suited to code generation (you can inline and even vectorize expressions across rows, without hopping pointers in between rows). Postgres does not yet include a column store, but this piece of infrastructure is a big step in the right direction!
It'll definitely not help for that, at least for now. There's not yet any caching around JITed programs, which makes it completely unsuitable for OLTP. With caching (which requires some larger changes to postgres, which is why this won't be in v11), it's possible to observe some moderate gains.
> I wonder if AWS will incorporate these improvements into Redshift, which is Postgres forked for OLAP use cases where this is far more prevalent. In the past, they haven't done a great job at this (ran into a float bug fixed upstream 5+ years ago), but this may be a big enough value proposition to justify the effort.
I can't imagine them doing that. This uses a whole rewritten expression evaluation engine (introduced in v10), and a large number of infrastructure changes all over the executor (reimplementation of targetlist set returning functions, projection, ...). Not impossible, but certainly not something lightly done.
There's not yet any caching around JITed programs,
which makes it completely unsuitable for OLTP.
I can't imagine them doing that.
--
This is a step in the right direction regardless, but I'm curious how many people are specifically using pure Postgres for large-scale OLAP models (not including occasionally running one-off queries on their OLTP data).
It has been merged into the master branch, which'll become PostgreSQL 11. First beta will be in a few weeks. Release is planned to be in September (early I'd hope).
> Will it be behind a flag, or does the planner just know when to use it?
There's cost based logic that chooses it, yes. See https://www.postgresql.org/docs/devel/static/jit-decision.ht...
You can change those if you want to force JIT on/off.
I know the RDS team has contributed some fixes and the like to Postgres, but I’m unsure of their overall significance.
I wonder if AWS will incorporate these improvements into Redshift, which is Postgres forked for OLAP use cases where this is far more prevalent. In the past, they haven't done a great job at this (ran into a float bug fixed upstream 5+ years ago), but this may be a big enough value proposition to justify the effort.