Describe polling with ETag caching

CONTRIBUTING.md

@@ -403,8 +403,8 @@ There are a few rules to get your merge request accepted:
    - Avoid repeated polling of endpoints that require a significant amount of overhead
    - Check for N+1 queries via the SQL log or [`QueryRecorder`](https://docs.gitlab.com/ce/development/merge_request_performance_guidelines.html)
    - Avoid repeated access of filesystem
-1. If you need polling to support real-time features, consider using this [described long
-   polling approach](https://gitlab.com/gitlab-org/gitlab-ce/issues/26926).
+1. If you need polling to support real-time features, please use
+   [polling with ETag caching][polling-etag].
 1. Changes after submitting the merge request should be in separate commits
    (no squashing). If necessary, you will be asked to squash when the review is
    over, before merging.
@@ -547,6 +547,7 @@ available at [http://contributor-covenant.org/version/1/1/0/](http://contributor
 [UX Guide for GitLab]: http://docs.gitlab.com/ce/development/ux_guide/
 [license-finder-doc]: doc/development/licensing.md
 [GitLab Inc engineering workflow]: https://about.gitlab.com/handbook/engineering/workflow/#labelling-issues
+[polling-etag]: https://docs.gitlab.com/ce/development/polling.html
 
 [^1]: Specs other than JavaScript specs are considered backend code. Haml
       changes are considered backend code if they include Ruby code other than just

doc/development/polling.md

@@ -0,0 +1,41 @@
+# Polling with ETag caching
+
+Polling for changes (repeatedly asking server if there are any new changes)
+introduces high load on a GitLab instance, because it usually requires
+executing at least a few SQL queries. This makes scaling large GitLab
+instances (like GitLab.com) very difficult so we do not allow adding new
+features that require polling and hit the database.
+
+Instead you should use polling mechanism with ETag caching in Redis.
+
+## How to use it
+
+1. Add the path of the endpoint which you want to poll to
+   `Gitlab::EtagCaching::Middleware`.
+1. Implement cache invalidation for the path of your endpoint using
+   `Gitlab::EtagCaching::Store`. Whenever a resource changes you
+   have to invalidate the ETag for the path that depends on this
+   resource.
+1. Check that the mechanism works:
+   - requests should return status code 304
+   - there should be no SQL queries logged in `log/development.log`
+
+## How it works
+
+1. Whenever a resource changes we generate a random value and store it in
+   Redis.
+1. When a client makes a request we set the `ETag` response header to the value
+   from Redis.
+1. The client caches the response (client-side caching) and sends the ETag as
+   the `If-None-Modified` header with every subsequent request for the same
+   resource.
+1. If the `If-None-Modified` header matches the current value in Redis we know
+   that the resource did not change so we can send 304 response immediately,
+   without querying the database at all. The client's browser will use the
+   cached response.
+1. If the `If-None-Modified` header does not match the current value in Redis
+   we have to generate a new response, because the resource changed.
+
+For more information see:
+- [RFC 7232](https://tools.ietf.org/html/rfc7232)
+- [ETag proposal](https://gitlab.com/gitlab-org/gitlab-ce/issues/26926)