Boost performance by batching Rows

[fabianfett]

In #135 a connection state machine was introduced. To keep this simple at first, the state machine operates on single PostgresBackendMessages. This is very inefficient for queries that result in a lot of rows. This PR improves query performance by badging Postgres data row messages at the Decoder and operating on those batches from there on.

Open ends:

• Two open warnings

[fabianfett]

Reduction in Actions 🥳

[fabianfett]

@weissi This is the invocation, where the PSQLRowStream asks for more rows. We could change this to:

context.channel.triggerUserOutboundEvent(PSQLRowStreamEvent.requestMoreRows, promise: nil)

[fabianfett]

Adding some performance numbers... (Swift main on focal in docker)

Running a query SELECT * FROM "large_table" that returns 1million rows with three number each:

1.4.4 release (before state machine):

root@3bcc272414db:/vapor/postgres-perf# time .build/release/PostgresPerf
2021-05-02T13:29:45+0000 info psql : PostgresQueryMetadata(command: "SELECT", oid: nil, rows: Optional(1000000))
2021-05-02T13:29:45+0000 info psql : took: 3.9132070541381836 seconds

real	0m4.455s
user	0m1.970s
sys	0m0.376s

main branch with state machine (perf regression):

root@3bcc272414db:/vapor/postgres-perf# time .build/release/PostgresPerf
2021-05-02T13:12:26+0000 info psql : PostgresQueryMetadata(command: "SELECT", oid: nil, rows: Optional(1000000))
2021-05-02T13:12:26+0000 info psql : query took: 5.985878944396973 seconds

real	0m6.256s
user	0m5.955s
sys	0m0.135s

this branch:

root@3bcc272414db:/vapor/postgres-perf# time .build/release/PostgresPerf
2021-05-02T13:13:23+0000 info psql : PostgresQueryMetadata(command: "SELECT", oid: nil, rows: Optional(1000000))
2021-05-02T13:13:23+0000 info psql : query took: 3.5723689794540405 seconds

real	0m4.003s
user	0m1.110s
sys	0m0.435s

While the speed bump is nice. Even more important: The query doesn't use 100% cpu anymore.

[jdmcd]

This is really enticing, any chance we can port this easily to MySQL as well? 👀

[fabianfett]

@jdmcd Since this pr builds on top of #135, I'm afraid some groundwork would need to be done to have this land.

[jdmcd]

Ah, I see

[gwynne]

This is really enticing, any chance we can port this easily to MySQL as well? 👀

@jdmcd FWIW this is something I've already got on my list to look into for MySQL 🙂

[fabianfett]

Just applied another fix. Will push shortly. On Swift 5.5 nightly:

Read 1m rows into an array
main        : 5.354 sec
this branch : 1.790 sec

Consume 1m rows on el
main        : 5.176 sec
this branch : 1.475 sec