Action API should permit CDN cache when generator is used (MobileFrontend search is slow)
Open, Needs TriagePublic
Actions

Assigned To

None

Authored By

	• Krinkle
	Jan 25 2022, 9:57 PM

Description

I'm noticing numerous common API queries, including those used by search suggestions on MobileFrontend, to not enjoy any form of CDN caching.

View https://en.m.wikipedia.org/wiki/Main_Page
Open network devtools.
Type letter "B".

https://en.m.wikipedia.org/w/api.php?action=query&format=json&formatversion=2&prop=pageprops%7Cpageprops%7Cpageimages%7Cdescription&generator=prefixsearch&ppprop=displaytitle&piprop=thumbnail&pithumbsize=80&pilimit=15&redirects=&gpssearch=B&gpsnamespace=0&gpslimit=15

Latency is 254ms (!)

Cache headers

cache-control: private, must-revalidate, max-age=0
x-cache: cp3050 miss, cp3058 pass
x-cache-status: pass

Compared to the equivalent query issued by Vector:
https://en.wikipedia.org/w/api.php?action=opensearch&format=json&formatversion=2&search=B&namespace=0&limit=10

Latency is just 11ms.

Cache headers

cache-control: max-age=10800, s-maxage=10800, public
x-cache: cp3058 miss, cp3058 hit/135
x-cache-status: hit-front

To narrow this down, I have confirmed this to be the case even when:

client is logged out and has no session (e.g. private browsing), although CDN cache should work for search suggestions even when logged-in (like it always has for OpenSearch as used by Vector, etc.).
when checking appserver response directly, it's not a bug in Varnish or other routing/traffic layers.
when querying without any of the more complex extensions like PageImages etc. Even a plain prefixsearch with generator enabled for basic pageinfo, reproduces the same bug locally.

 $ curl -i 'http://localhost:8080/w/api.php?action=query&format=json&formatversion=2&generator=prefixsearch&gpssearch=B' 
HTTP/1.1 200 OK
Server: Apache/2.4.38 (Debian)
Cache-Control: private, must-revalidate, max-age=0
..
{"batchcomplete":true,"query":{"pages":[{"pageid":8,"ns":0,"title":"Bar","index":1}]}}

 $ curl -i 'http://localhost:8080/w/api.php?action=opensearch&format=json&formatversion=2&search=B&namespace=0&limit=10'
HTTP/1.1 200 OK
Server: Apache/2.4.38 (Debian)
Cache-Control: max-age=1200, s-maxage=1200, public
..
["B",["Bar"],[""],["http://mw.localhost:8080/wiki/Bar"]]

Related Objects

Mentioned Here: T97096: API requests are not cached for logged-in users unless uselang is set explicitly

Event Timeline

• Krinkle created this task.Jan 25 2022, 9:57 PM

Restricted Application added a subscriber: Aklapper. · View Herald TranscriptJan 25 2022, 9:57 PM

I added some rudimentary instrumentation to narrow down where it might be going wrong:

--- a/includes/api/ApiMain.php
+++ b/includes/api/ApiMain.php
@@ -793,6 +793,7 @@ class ApiMain extends ApiBase {
        }
 
        wfDebug( __METHOD__ . ": setting cache mode $mode" );
+       header( "X-Log-cachemode-4-set: " . $mode );
        $this->mCacheMode = $mode;
    }
 
diff --git a/includes/api/ApiQuery.php b/includes/api/ApiQuery.php
index dc5b4d8b8cd..8bc18b7c1ee 100644
--- a/includes/api/ApiQuery.php
+++ b/includes/api/ApiQuery.php
@@ -620,10 +620,12 @@ class ApiQuery extends ApiBase {
        }
 
        $cacheMode = $this->mPageSet->getCacheMode();
+       header( 'X-Log-cachemode-1-pageset: ' . $cacheMode );
 
        // Execute all unfinished modules
-       foreach ( $modules as $module ) {
+       foreach ( $modules as $k => $module ) {
            $params = $module->extractRequestParams();
+           header( "X-Log-cachemode-2-$k-" . get_class($module) . ': ' . $module->getCacheMode( $params ) );
            $cacheMode = $this->mergeCacheMode(
                $cacheMode, $module->getCacheMode( $params ) );
            $module->execute();
@@ -631,6 +633,7 @@ class ApiQuery extends ApiBase {
        }
 
        // Set the cache mode
+       header( "X-Log-cachemode-3-set: " . $cacheMode );
        $this->getMain()->setCacheMode( $cacheMode );
 
        // Write the continuation data into the result

This yields:

curl -i 'http://localhost:8080/w/api.php?action=opensearch&format=json&formatversion=2&search=B&namespace=0&limit=10'
X-Log-cachemode-4-set: anon-public-user-private
Cache-Control: max-age=1200, s-maxage=1200, public

curl -i 'http://localhost:8080/w/api.php?action=query&format=json&formatversion=2&generator=prefixsearch&gpssearch=B&gpsnamespace=0&gpslimit=15' 
X-Log-cachemode-1-pageset: public
X-Log-cachemode-3-set: public
X-Log-cachemode-4-set: anon-public-user-private
Cache-Control: private, must-revalidate, max-age=0

$ curl -i 'http://localhost:8080/w/api.php?action=query&format=json&formatversion=2&generator=prefixsearch&gpssearch=B' 
X-Log-cachemode-1-pageset: public
X-Log-cachemode-3-set: public
X-Log-cachemode-4-set: anon-public-user-private
Cache-Control: private, must-revalidate, max-age=0

That rules out pretty much all middle layers where I suspected the cause would be.

• Krinkle updated the task description. (Show Details)Jan 25 2022, 10:27 PM

• Krinkle edited projects, added MediaWiki-Action-API; removed MediaWiki-REST-API.

• Krinkle edited projects, added Platform Engineering, Performance-Team (Radar); removed Performance-Team.

• Krinkle moved this task from Limbo to Perf recommendation on the Performance-Team (Radar) board.

Aklapper renamed this task from Action API should permit CDN cachje when generator is used (MobileFrontend search is slow) to Action API should permit CDN cache when generator is used (MobileFrontend search is slow).Jan 26 2022, 1:57 AM

@Jdlrobson The above may be impacting mobile UX. Feel free to radar/track away, this is just an FYI. Though if you happen to remember a time when this wasn't slow/uncached, that'd help narrow it down.

Thanks for the ping. I am not aware of any changes that might have led to this slow down. Frontend hasn't been touched since 2018. Team doesn't work on backend, so perhaps some changes have occurred there?

@ovasileva this might be good motivation to prioritize the work on mobile to switch it to the new search widget.

Jdlrobson moved this task from Backlog to Using Codex in web team products / Skin unification on the Web-Team-Backlog-Archived (Needs Prioritization (Tech)) board.Feb 3 2022, 8:13 PM

In T300078#7676652, @Jdlrobson wrote:

Thanks for the ping. I am not aware of any changes that might have led to this slow down. Frontend hasn't been touched since 2018. Team doesn't work on backend, so perhaps some changes have occurred there?

@ovasileva this might be good motivation to prioritize the work on mobile to switch it to the new search widget.

@Jdlrobson - do we know how the UX/frontend might be affected here?

If the API is slower, then searching will feel slower and it will take longer for results to display.

Jdlrobson edited projects, added Web-Team-Backlog-Archived; removed Web-Team-Backlog-Archived (Needs Prioritization (Tech)).Mar 15 2023, 10:37 PM

Jdlrobson moved this task from Incoming to Tracking on the Web-Team-Backlog-Archived board.Mar 21 2023, 5:11 PM

• Krinkle edited projects, added Wikimedia-Performance-recommendation; removed Performance-Team (Radar).Aug 18 2023, 8:41 PM

What does it mean to cache a generator query response? The next time you call it, chances are the generated page set will be different.

The action API mostly leaves the decision to cache or not cache with the client; when you use maxage/smaxage, the result will be cacheable:

curl -i '<wiki>/w/api.php?action=query&format=json&formatversion=2&generator=prefixsearch&gpssearch=B&maxage=100&smaxage=100' 
HTTP/1.1 200 OK
Cache-control: s-maxage=100, max-age=100, public
Expires: Tue, 04 Feb 2025 08:27:47 GMT

although CDN cache should work for search suggestions even when logged-in

See T97096: API requests are not cached for logged-in users unless uselang is set explicitly about that.

Action API should permit CDN cache when generator is used (MobileFrontend search is slow)Open, Needs TriagePublicActions

Description

Related Objects

Event Timeline

Action API should permit CDN cache when generator is used (MobileFrontend search is slow)
Open, Needs TriagePublic
Actions