Findings documented in https://docs.google.com/document/d/1lGF8X00_8NhGyqkOeNDJFgq9eFW375JER13HLQHBp-Q

Pulling from the doc above, we saw roughly 1.3M errors over the sampled week May 13-20, and of those 1.1M were the "groupIds not available" message.

Spot-checking the huge number of bad groupIds, I'm not coming up with a general theory yet. To take one example, the article for the Russian Embassy in Bejing on Chinese Wikipedia is requested roughly a dozen times per day, with a consistent but incorrect map hash. This map is coming from a template and pulling the Wikidata coordinate for the page item, but neither the coordinate nor the template seem to have changed within the last few months. Ultimately, the mapframe is generated by this Lua module which hasn't changed in years.

What's even more confounding is that the map image request is now being generated with revids for versioning, (sample link), and this revid is for the latest revision. The issue can be seen before versioned maps were deployed, with the same incorrect groupId being requested.

I'll check the referrer and agent for these requests, to be sure these are regular users visiting the page and not bots or coming from an off-wiki source. But the revids parameter is a strong indicator that MediaWiki is at fault.

Interesting lead to follow up on: the lang parameter to Kartotherian is always zh, however there seems to be a relationship between the bad map group ID and the request's accept-language being "zh-cn". Indeed, setting uselang on the page view https://zh.wikipedia.org/wiki/%E4%BF%84%E7%BE%85%E6%96%AF%E9%A7%90%E8%8F%AF%E5%A4%A7%E4%BD%BF%E9%A4%A8?uselang=zh-cn results in the broken map: https://maps.wikimedia.org/img/osm-intl,10,a,a,270x200.png?lang=zh&domain=zh.wikipedia.org&title=%E4%BF%84%E7%BE%85%E6%96%AF%E9%A7%90%E8%8F%AF%E5%A4%A7%E4%BD%BF%E9%A4%A8&revid=71713591&groups=_bc660fd4ba69fba1a1a083ce694bb530f37ad88f

Requesting the mapdata with a uselang=zh-cn parameter results in the bad group ID:
https://zh.wikipedia.org/w/api.php?action=query&prop=mapdata&revids=60731942&uselang=zh-cn

This suggests that parser cache is being split by interface language, within the mapframe. Here's the livedata with uselang=en,

{
  "_dc1eec1fc2c94a5130dac46def8fec9060c3fdfb": [
    {
      "type": "ExternalData",
      "service": "geoshape",
      "url": "https://maps.wikimedia.org/geoshape?getgeojson=1&ids=Q4374043",
      "properties": {
        "stroke-width": 3,
        "stroke": "#FF0000",
        "title": "俄羅斯駐華大使館",
        "fill": "#606060"
      }
    },
    {
      "type": "ExternalData",
      "service": "geoline",
      "url": "https://maps.wikimedia.org/geoline?getgeojson=1&ids=Q4374043",
      "properties": {
        "stroke-width": 5,
        "stroke": "#FF0000",
        "title": "俄羅斯駐華大使館"
      }
    },
    {
      "type": "Feature",
      "geometry": {
        "coordinates": [
          116.422897,
          39.944452
        ],
        "type": "Point"
      },
      "properties": {
        "title": "俄羅斯駐華大使館",
        "marker-color": "#5E74F3"
      }
    }
  ]
}

and with uselang=zh-cn,

{
  "_bc660fd4ba69fba1a1a083ce694bb530f37ad88f": [
    {
      "type": "ExternalData",
      "service": "geoshape",
      "url": "https://maps.wikimedia.org/geoshape?getgeojson=1&ids=Q4374043",
      "properties": {
        "stroke-width": 3,
        "stroke": "#FF0000",
        "title": "俄罗斯驻华大使馆",
        "fill": "#606060"
      }
    },
    {
      "type": "ExternalData",
      "service": "geoline",
      "url": "https://maps.wikimedia.org/geoline?getgeojson=1&ids=Q4374043",
      "properties": {
        "stroke-width": 5,
        "stroke": "#FF0000",
        "title": "俄罗斯驻华大使馆"
      }
    },
    {
      "type": "Feature",
      "geometry": {
        "coordinates": [
          116.422897,
          39.944452
        ],
        "type": "Point"
      },
      "properties": {
        "title": "俄罗斯驻华大使馆",
        "marker-color": "#5E74F3"
      }
    }
  ]
}

It's traditional vs. simplified chinese, "華" vs. "华", so probably due to some exciting layer which intentionally splits by interface language preferences, regardless of the page's content language.

Note: geoshapes service errors need to have additional fields added to the log messages, so that we can connect them to a specific page.