Our setup is a bit complex, so I will try to simplify what our tables look like but still try to describe the real problem:
We have a large table (
items) with about 500 million rows (grows with ~1.5million rows per day). Where each item has a timestamp (among other things).
We have 5 metadata tables (based on the item type) with FKs to
items. These have about 100 million rows each.
When doing queries on the
items table, we need to filter and sort on values in the metadata tables. This is based on filter parameters in our UI, so we might filter on things from all 5 metadata tables, or from none of them. The query will look something like this:
SELECT i.id FROM items i LEFT JOIN itemmetadata1 im1 ON im1.itemid = i.id AND i.type = 1 LEFT JOIN itemmetadata2 im1 ON im2.itemid = i.id AND i.type = 2 LEFT JOIN itemmetadata3 im1 ON im3.itemid = i.id AND i.type = 3 LEFT JOIN itemmetadata4 im1 ON im4.itemid = i.id AND i.type = 4 LEFT JOIN itemmetadata5 im1 ON im5.itemid = i.id AND i.type = 5 WHERE i.timestamp > '2019-02-01' AND i.timestamp < '2019-02-07' -- These aren't always here, query is dynamically generated (based on user input) AND im1.somevalue = TRUE AND im3.anothervalue > 5 ... etc ORDER BY -- This can also change dynamically im5.value DESC
This makes this query quite slow for some cases. For example when looking at a long period of time, with few matching rows in the metadata tables.
So I have two questions:
Would it make sense to create a new denormalized table called
aggregateditems that contains all of the item and metadata columns we need to filter and sort on? That way our query would be simplified like this:
SELECT i.id FROM aggregateditems ai WHERE ai.timestamp > '2019-02-01' AND ai.timestamp < '2019-02-07' -- These aren't always here, query is dynamically generated (based on user input) AND ai.metadata1somevalue = TRUE AND ai.metadata3anothervalue > 5 ... etc ORDER BY -- This can also change dynamically ai.metadata5value DESC
I guess this would speed things up considerably with the right indexes.
So my thought is to put triggers on
items and the metadata tables, that will update
aggregateditems when the data changes, is that a good idea? Or might this kill performance?
Creating an aggregated table will speed up queries I think, but we'd still have very large tables, with large indexes, so I guess we'd need to do some partitioning as well.
Let's say we partition
timestamp, how do we partion the metadata tables? (They don't have any timestamps).