there's no need to cast the int array into a pg type array. Seems pgx
does the work for us!

Signed-off-by: Taylor Silva <dev@taydev.net>

had to create our own implementation of pq's listener. Thankfully we had
that wrapped in an interface already.

While digging into the pq implementation I found we couldn't implement
things in the exact same way they did. They actually received
notifications by using a sort of "escape hatch" they built into their
driver. They had a func that would read every single bit of data sent to
the driver. It would then check the type of the data going through the
driver and if it was a "Notification" then it would read the raw data
from a bufffer in the driver. We cannot do the same thing with the Pgx
driver.

I'm not 100% sure my implementation is good. It's passing most tests,
but I'm not sure how robust it is. I have no clue if pgx will recover
from connection errors for us for example.

Signed-off-by: Taylor Silva <dev@taydev.net>

This interface is embedded in the LogConn struct. Because we expect the
interface to have a method named "Conn" this resulted in a naming
conflict when the interface is embedded because the method name now
conflicts with the embedded name of the interface... yeah that's
confusing to explain. Anyways, had to rename Conn so LogConn could
satisfy the Conn interface, that's the gist of this commit.

Signed-off-by: Taylor Silva <dev@taydev.net>

Signed-off-by: Taylor Silva <dev@taydev.net>

Signed-off-by: Taylor Silva <dev@taydev.net>

Signed-off-by: Taylor Silva <dev@taydev.net>

Signed-off-by: Taylor Silva <dev@taydev.net>

and further decrease sleep time

Signed-off-by: Taylor Silva <dev@taydev.net>

Signed-off-by: Taylor Silva <dev@taydev.net>

Signed-off-by: Taylor Silva <dev@taydev.net>

using a pool so we can easily reset the connection if needed. the lib/pq
listener had this behaviour. We're just trying to replicate it here.
Using exponential backoff which will timeout after 15mins. We use this
package in other places in our codebase.

Signed-off-by: Taylor Silva <dev@taydev.net>

Signed-off-by: Taylor Silva <dev@taydev.net>

Signed-off-by: Taylor Silva <dev@taydev.net>

in this case we had row.BuildID with a type of int. The column we were
trying to insert that into was of type "text" in postgres. We have to
convert the int to a string first before passing it into the query.

Signed-off-by: Taylor Silva <dev@taydev.net>

WOW is the pgx driver ANAL about types when it comes to arrays.

So in postgres, if `array_agg()` returns nothing, you get back `{NULL}`.
So this _should_ cast just fine to the sql.NullString type from the Go
sql package. Pgx cannot figure out how to cast to this type at all, you
just get errors like this:

sql: Scan error on column index 3, name "passed": cannot convert to sql.Scanner: cannot find registered type for *[]sql.NullString

and initially I was getting this error:

sql: Scan error on column index 3, name "passed": unsupported Scan, storing driver.Value type string into type *[]sql.NullString

In that second error I found it weird that it said it was trying to turn
a type `string` into type `*[]sql.NullString`. I kept playing around
with the types and using dlv to debug. I eventually figured out pgx
would safely scan into a slice of type `[]*string`, so an array that
contains pointers to strings, which is a VERY annoying array to use!

I eventually figured out this is because `array_agg()` returns `{NULL}`
when it has no results. I figured out that by wrapping `array_agg()`
with `array_remove(array_agg(), NULL)` I could then get pgx to cast to a
normal `[]string` slice.

A bunch of tests still failed a this point because `Passed` on the
various structs would be non-nil. This is why there are length checks on
the slices and we set the var to `nil` if the length is zero before we
pass it to the returning struct.

I think this is actually a bit of a bug in pgx. This workaround is not
bad, but was VERY annoying to figure out, especially when it worked just
fine with lib/pq.

Signed-off-by: Taylor Silva <dev@taydev.net>

Signed-off-by: Taylor Silva <dev@taydev.net>

Signed-off-by: Taylor Silva <dev@taydev.net>

Signed-off-by: Taylor Silva <dev@taydev.net>

Signed-off-by: Taylor Silva <dev@taydev.net>

also learned that pgx does not support sending multiple commands in a
single statement. You have to send it in two separate statements:

jackc/pgx#1090

Signed-off-by: Taylor Silva <dev@taydev.net>

It seems the pgx driver gives up on the initial connection very quickly.
Guessing lib/pq made some re-attempts which is why we never saw this
behaviour previously.

Signed-off-by: Mathias Bogaert <mathias.bogaert@gmail.com>
Co-authored--by: Taylor Silva <dev@taydev.net>

Disables fsync, full_page_writes and synchronous_commit for
Postgres in CI environments. This greatly improves performance
by removing disk durability guarantees which aren't needed
in ephemeral test environments.

Sets random_page_cost=1.1 which tells PostgreSQL's query
planner that random disk access is relatively fast, as is
typical with SSDs. This encourages the planner to consider
index scans over sequential scans when appropriate.

Sets random_page_cost=1.1 which tells PostgreSQL's query
planner that random disk access is relatively fast, as is
typical with SSDs. This encourages the planner to consider
index scans over sequential scans when appropriate.

…emoved-async-testing

there was a race condition here where we would try listening for
notifications with a nil conn. Had to make sure the check on the
connection was AFTER we were given back control from the comms channel.
Also made it explicit that l.conn is set to nil during close() even
though it appears Release() does this for us. This is to avoid any race
conditions with Release()

Signed-off-by: Taylor Silva <dev@taydev.net>