Polling for patchseries in DPDK - the /series/ and /events/ endpoints

Polling for patchseries in DPDK - the /series/ and /events/ endpoints

[Patrick Robb]

[-- Attachment #1: Type: text/plain, Size: 4492 bytes --]

There was some discussion at last week's CI meeting about usage of the
Patchwork /events/ endpoint for polling for patches, and issues with that
process. Here is a relevant blurb, explaining some issues Aaron has run
into using the dpdk-ci repo "poll-pw.sh" shell script:

----------------

* Discussion pertaining to looking at polling for series using the events
API. This events endpoint (with series created event) returns info that a
series has been created, but returns a limited set of data in the payload,
and this necessitates a followup request to patchwork. So, this seems like
it would actually increase the amount of requests made to the patchwork
server. Some related issues discussed are:
   * You cannot query the events endpoint for only events from a particular
project (this matters for patchwork instances with many projects under
them). For DPDK there are only 4 projects under DPDK patchwork, so it’s not
a huge deal, but still a small issue.
   * The datetime that the series-created event returns is the datetimes of
one of the commits in the series, not the datetime of when the series was
submitted. So, this means that if you amend a commit (this does not update
commit datetime) and resubmit a patchseries, the datetime on the
series-created record will not be “updated”. This can cause us to miss
series when polling via the events endpoint.

------------------

And for context, poll-pw.sh will check the /events/ endpoint for new series
created events like so:

--------------------

URL="${URL}/events/?category=${resource_type}-completed"

callcmd () # <patchwork id>
{
	eval $cmd
}

while true ; do
	date_now=$(date --utc '+%FT%T')
	since=$(date --utc '+%FT%T' -d $(cat $since_file | tr '\n' ' '))
	page=1
	while true ; do
		ids=$(curl -s "${URL}&page=${page}&since=${since}" |
			jq "try ( .[] | select( .project.name == \"$project\" ) )" |
			jq "try ( .payload.${resource_type}.id )")
		[ -z "$(echo $ids | tr -d '\n')" ] && break
		for id in $ids ; do
			if grep -q "^${id}$" $poll_pw_ids_file ; then
				continue
			fi
			callcmd $id
			echo $id >>$poll_pw_ids_file


-------------------

But, as was discussed at the meeting, once you have the series ids, then
you need to make a followup request to /series/{id}.

UNH has a download_patchset.py polling script very much like poll-pw.sh
except that, because we store extra info about our processed patchseries in
a database (to facilitate lab.dpdk.org filtering functions), we use our
database to get the most recently processed patchseries, instead of the
"since_file." Our process (running every 10 minutes from Jenkins) is like
this:

1. get the "since_id" from our database
2. get the "newest_id" from https://patchwork.dpdk.org/api/
events/?category=series-completed. Get the [0] index of the json response
(the most recent patchseries) and save that series id.
3. for seriesID in range(since_id, newest_id): get patch from
https://patchwork.dpdk.org/api/series/{id}.

So, both poll-pw.sh and our UNH script follow the process of making a
request to /events/, and then followup requests for /series/. Thus the
total number of requests being made on patchwork is (number of new
patchseries + 1).

-The most consequential difference in the two implementations is that
poll-pw.sh makes a request to /events/ with the &since=${since} parameter,
passing in a since datetime, and UNH does not. As Aaron explained at the CI
meeting, because the datetime provided in the /events/ payload is not what
one would expect (it gives the datetime of the commit, not when the series
was submitted) this means that poll-pw-sh can miss series. With the UNH lab
polling script we don't have this issue because we don't make use of the
since parameter in our /events/ request. I think the options for poll-pw.sh
going forward would be:
1. Update patchwork so that the datetime provided in the /events/ payload
is what is "expected" i.e. the datetime that the series was submitted at.
2. Adopt the UNH process of discarding the &since=${since} parameter, and
rely solely on tracking the most recently processed patchseries id, get the
newest patchseries id from /events/, and traverse the range of (since_id,
newest_id).

-I agree it makes sense for /events/ to support a "project" param.

Thanks Aaron for raising this conversation. We can continue the
conversation over email, or also in person at DPDK Prague!

[-- Attachment #2: Type: text/html, Size: 5503 bytes --]