Implement encode/decode for ID columns

[calebj]

This PR allows users to specify the special value of 'func' for p_epoch to use custom functions to encode/decode time-ordered integers other than the classic seconds, ms, us or ns since the UNIX epoch.

Resolves #729.

[calebj]

I removed this variable since it isn't used anywhere.

[calebj]

This should remove the need for a functional index to find the maximum, provided there is an index on the underlying column. That should be a safe assumption to make since postgres prohibits creating a unique index or primary key on a partitioned table that doesn't contain the partition key.

If this works in all scenarios, it should also be ported to partition_data_time()

[calebj]

Marking as draft because this still needs tests for smaller integer types

[keithf4]

Thank you! Will be a bit busy over the next month or so, so may not get to dig into this for a bit.

Just reading your notes, though, it is possible to have unique indexes on non-control columns for individual tables using the template table system in partman. May also want to keep in mind that it may be possible in core in the future as well. So not sure if that will affect what you've done.

[calebj]

Unique indexes on individual partitions or on non-control columns won't affect anything in this PR. The intent is to optimize the check by using MAX() on the base column before converting it to a timestamp, since otherwise it needs a functional index. I added that note because if there's no edge case where that would be undesirable, it makes sense to apply the same type of change to this section's calls to MIN() and MAX()

pg_partman/sql/functions/partition_data_time.sql

Lines 131 to 159 in 1192e93

	v_partition_expression := CASE
	WHEN v_epoch = 'seconds' THEN format('to_timestamp(%I)', v_control)
	WHEN v_epoch = 'milliseconds' THEN format('to_timestamp((%I/1000)::float)', v_control)
	WHEN v_epoch = 'microseconds' THEN format('to_timestamp((%I/1000000)::float)', v_control)
	WHEN v_epoch = 'nanoseconds' THEN format('to_timestamp((%I/1000000000)::float)', v_control)
	ELSE format('%I', v_control)
	END;
	-- Generate column list to use in SELECT/INSERT statements below. Allows for exclusion of GENERATED (or any other desired) columns.
	SELECT string_agg(quote_ident(attname), ',')
	INTO v_column_list
	FROM pg_catalog.pg_attribute a
	JOIN pg_catalog.pg_class c ON a.attrelid = c.oid
	JOIN pg_catalog.pg_namespace n ON c.relnamespace = n.oid
	WHERE n.nspname = v_source_schemaname
	AND c.relname = v_source_tablename
	AND a.attnum > 0
	AND a.attisdropped = false
	AND attname <> ALL(COALESCE(p_ignored_columns, ARRAY[]::text[]));
	FOR i IN 1..p_batch_count LOOP
	IF p_order = 'ASC' THEN
	EXECUTE format('SELECT min(%s) FROM ONLY %I.%I', v_partition_expression, v_source_schemaname, v_source_tablename) INTO v_start_control;
	ELSIF p_order = 'DESC' THEN
	EXECUTE format('SELECT max(%s) FROM ONLY %I.%I', v_partition_expression, v_source_schemaname, v_source_tablename) INTO v_start_control;
	ELSE
	RAISE EXCEPTION 'Invalid value for p_order. Must be ASC or DESC';
	END IF;

One other improvement I want to make is a round-trip and ordering sanity check for the passed functions, and an assert that the functions are at least marked RETURNS NULL ON NULL INPUT or STRICT, and IMMUTABLE.

Also, I think Julian dates might make an interesting example for 32-bit IDs, so I will play with that for a daily test.