[sgarland]

I assumed that they were storing the ULIDs as binary, in the UUID column type, as link 2 in your reply. If stored as TEXT, then yes, that absolutely would make a difference.

It’s also worth noting that unlike MySQL / SQL Server, Postgres does not store tuples clustered around the PK. Indices are of course still in a B+tree.

[atombender]

They show that they're storing the ULIDs as text. Quoting from the article:

   CREATE TABLE ulid_test(id TEXT);

I suspect their poor results come from their choice of ULID implementation. The native C implementation I tried out is faster than the Postgres UUID type when testing computation only.

I noticed a bug in their test: They call generate_ulid() with now(). But now() is an alias for transaction_timestamp(), which is computed once at the start of the transaction, so all the timestamps will be the same. They should be using clock_timestamp().

[sgarland]

Good catch to both.