Rest UpdateCounter not returning correct values

[hloeffler]

What happened:
api is returning wrong data for UpdateCounter

What you expected to happen:
Get current count and capacity
but got:

{"name":"","count":"0","capacity":"0"}

How to reproduce it (as minimally and precisely as possible):

curl -sSd '{"count": "5'", "capacity": "64"}' -H "Content-Type: application/json" -X PATCH http://localhost:${AGONES_SDK_HTTP_PORT}/v1beta1/counters/players

Anything else we need to know?:
both return the right data:

curl http://localhost:${AGONES_SDK_HTTP_PORT}/v1beta1/counters/players

->

{"name":"players","count":"39","capacity":"64"}

curl http://localhost:${AGONES_SDK_HTTP_PORT}/gameserver

->

    "ports": [
      {
        "name": "test",
        "port": 13743
      }
    ],
    "players": null,
    "counters": {
      "players": {
        "count": "58",
        "capacity": "64"
      }
    },
    "lists": {}

Environment:

• Agones version: 1.49.0

[hloeffler]

was expecting behavior according to docs: https://agones.dev/site/docs/guides/client-sdks/rest/

[markmandel]

GetGameServer() will return whatever is latest that has been pushed up to the SDK Server from the CRD, Counter and List operations are atomic to local changes.

We should change this line in the docs:

The SDK batches mutation operations every 1 second for performance reasons. However, changes made and subsequently retrieved through the SDK will be atomically accurate through the SDK, as those values are tracked within the SDK Server sidecar process.

to something akin to (emphasis on change)
https://agones.dev/site/docs/guides/client-sdks/#counters-and-lists

The SDK batches mutation operations every 1 second for performance reasons. However, changes made and subsequently retrieved through the counter and list SDK functions will be atomically accurate through the SDK, as those values are tracked within the SDK Server sidecar process.

Since it does seem to cause confusion! 😄

[hloeffler]

thanks for your fast response @markmandel

i'm still confused as my "gameserver code" looks like:

for i in {1..10}; do
  echo "update ccu ($i)"
  sleep 10
  PLAYERS=$((RANDOM % 64))
  curl -sSd '{"count": "'"${PLAYERS}"'", "capacity": "64"}' -H "Content-Type: application/json" -X PATCH http://localhost:${AGONES_SDK_HTTP_PORT}/v1beta1/counters/players
done

and the output is:

{"name":"","count":"0","capacity":"0"}update ccu (2)
{"name":"","count":"0","capacity":"0"}update ccu (3)
{"name":"","count":"0","capacity":"0"}update ccu (4)
{"name":"","count":"0","capacity":"0"}update ccu (5)
{"name":"","count":"0","capacity":"0"}update ccu (6)

a curl:
curl http://localhost:${AGONES_SDK_HTTP_PORT}/v1beta1/counters/players
gives me the correct name and numbers:
{"name":"players","count":"42","capacity":"64"}

[markmandel]

Just to confirm something - can you output the whole JSON you get back from when you are retrieving the gameserver information as a whole from the SDK?

[hloeffler]

sure:
a

PLAYERS=$((RANDOM % 64))

curl -sSd '{"count": "'$PLAYERS'", "capacity": "64"}' -H "Content-Type: application/json" -X PATCH http://localhost:${AGONES_SDK_HTTP_PORT}/v1beta1/counters/players

curl -sSH "Content-Type: application/json" -X GET http://localhost:${AGONES_SDK_HTTP_PORT}/gameserver | jq '.'

gives me:

{"name":"","count":"0","capacity":"0"}
{
  "object_meta": {
    "name": "xxx",
    "namespace": "xxx",
    "uid": "xxx",
    "resource_version": "628828067",
    "generation": "4",
    "creation_timestamp": "1753127942",
    "deletion_timestamp": "0",
    "annotations": {
      "agones.dev/sdk-version": "1.49.0"
    },
    "labels": {
xxx
    }
  },
  "spec": {
    "health": {
      "disabled": true,
      "period_seconds": 0,
      "failure_threshold": 0,
      "initial_delay_seconds": 0
    }
  },
  "status": {
    "state": "Scheduled",
    "address": "xxx",
    "addresses": [
      {
        "type": "InternalIP",
        "address": "xxx"
      },
      {
        "type": "ExternalIP",
        "address": "xxx"
      },
      {
        "type": "Hostname",
        "address": "xxx"
      },
      {
        "type": "PodIP",
        "address": "xxx"
      }
    ],
    "ports": [
      {
        "name": "test",
        "port": 14271
      }
    ],
    "players": null,
    "counters": {
      "players": {
        "count": "0",
        "capacity": "0"
      }
    },
    "lists": {}
  }
}

[hloeffler]

when i add a sleep 5:

PLAYERS=$((RANDOM % 64))

curl -sSd '{"count": "'$PLAYERS'", "capacity": "64"}' -H "Content-Type: application/json" -X PATCH http://localhost:${AGONES_SDK_HTTP_PORT}/v1beta1/counters/players

curl -sSH "Content-Type: application/json" -X GET http://localhost:${AGONES_SDK_HTTP_PORT}/gameserver | jq '.'

sleep 5

curl -sSH "Content-Type: application/json" -X GET http://localhost:${AGONES_SDK_HTTP_PORT}/gameserver | jq '.'

i get:

...
    "ports": [
      {
        "name": "test",
        "port": 12073
      }
    ],
    "players": null,
    "counters": {
      "players": {
        "count": "60",
        "capacity": "64"
      }
    },
    "lists": {}
  }

[markmandel]

when i add a sleep 5:

Well yeah - if you're in a tight loop like that, and there's no room for the sdk sidecar to get new updates... well, it won't get new updates.

Sounds like this was an issue with you overloading the CPU?

[hloeffler]

@markmandel
sleep is bash and seconds, there is no load at all

[hloeffler]

even if there is overload why should name: be "" in the return ?

[markmandel]

even if there is overload why should name: be "" in the return ?

Because there's no room in the CPU cycles for the information about the gameserver to be populated after the change.

[hloeffler]

i use only the rest api in a sidecar
it has nothing to do for like 2 minutes besides an updating the player number every 10 seconds.

[markmandel]

So you'll note that SDK methods are eventually consistent. It batches update operations and also CRD updates have to be sent to the K8s controller and then come back to the SDK.

So if you do a counter update (or really any update through the SDK) it's quite likely it won't be immediately available to GetGameServer(), as it takes time to process.

See this section of the docs:

[markmandel]

I swear I did note this somewhere in an issue - the docs here are somewhat confusing - in that "the change made and subsequently retrieved through the counter and list SDK functions will be atomically accurate."

We should go make that change to the docs, because it is an easy thing to interpret as all SDK methods.