I have a dev team requesting that I partition 3 very large tables with more 2 billion rows each. I’ve already got the partition key decided for each with a script which will create a copy of original with the partition declaration, along with the partition child tables. The problem will be that each table has an API which updates each table fairly regularly, and the devs say it cannot be turned off. So, I’m looking for the best way to do this so that the partitioned version will match the original table. FYI, this is production instance that is regularly updated by field technicians. Quite literally, I cannot perform two sequential select count(*) from any of these tables which will return the same row counts. I’m thinking of installing a trigger on insert on the original tables, which would write any insert to a third table, then copy the original table to another table name, do the partitioning on it to yet another table name, move the partition table into the original table name, kill the trigger, then apply the updates from the table where the insert trigger was being applied. Is there a better way to do this, or am I missing something? — Jay
Sent from my iPad
Are you planning on using the built in, native partitioning or using the old method combining trigger, constraints and inheritance? That should be the first thing that's decided since it will greatly influence the method you use. I'd highly recommend going with the native, built-in if at all possible since it has significantly better write performance and better read performance on PG12+.
If you're using time or integer based partitioning, the pg_partman extension has a lot of utilities and methods to help manage and migrate to a partitioned table. Check out the howto documentation for some examples of doing the kind of live migrations similar to what you may be looking for.