Re: Documenting when to retry on serialization failure

Fwiw I think the real problem with automatic retries is that the SQL
interface doesn't lend itself to it because the server never really
knows if the command is going to be followed by a commit or more
commands.

I actually think if that problem were tackled it would very likely be
a highly appreciated option. Because I think there's a big overlap
between the set of users interested in higher isolation levels and the
set of users writing stored procedures defining their business logic.
They're both kind of "traditional" SQL engine approaches and both lend
themselves to the environment where you have a lot of programmers
working on a system and you're not able to do things like define
strict locking and update orderings.

So a lot of users are probably looking at something like "BEGIN;
SELECT create_customer_order(....); COMMIT" and wondering why the
server can't handle automatically retrying the query if they get an
isolation failure.

There are actually other reasons why providing the whole logic for the
transaction up front with a promise that it'll be the whole
transaction is attractive. E.g. vacuum could ignore a transaction if
it knows the transaction will never look at the table it's
processing... Or automatic deadlock testing tools could extract the
list of tables being accessed and suggest "lock table" commands to put
at the head of the transaction sorted in a canonical order.

These things may not be easy but they're currently impossible for the
same reasons automatically retrying is. The executor doesn't know what
subsequent commands will be coming after the current one and doesn't
know whether it has the whole transaction.