This site rendered js and css (now even images) dynamically. Contents of files are stored in Google BigTable Datastore and are served by a DynServlet. The DynServlet maps the name of the file to content of the file in DB. This was slowly using up a lot of resources. Adding a layer of memcache, saved some Datastore API calls. But the memcache calls were still significant. So an ETag mechanism is introduced.
A Global ETag is generated on any file posted or updated, and (G-ETag is) stored in Datastore. To save App Engine resources (which are counted to cost), the ETag header is obtained from each request checked with GDTS from backend. If it is the same a "304 – Not Modified" response is sent back, asking the client to use the local cache copy. If it is not exactly same, a the request is passed through and rendered dynamically. This guarantees all content will be served fresh after every deployment. Requests in between deployments may be served from client-side cache, because they will have same ETag.
However, having just ETag will still generate "GET " request and "304 – Not Modified" response (ping) conversation for each request, even if it has the latest version in the cache. Meaning, there will still be network round trips for each page. So a Cache-Control header is also used. This header, Cache-Control: public, max-age:3600. caches the content at the BrowserAgent (client) for an hour. So the change pings will not called at least for 1 hr.
This way, it is guaranteed that data can be stale at most 1 hr. (can also be considered downside). However, according to the site's analytics, max length of time spent by a user on the site is 14 mins. 1 hr. covers the most dedicated visitor :)
A Global ETag is generated on any file posted or updated, and (G-ETag is) stored in Datastore. To save App Engine resources (which are counted to cost), the ETag header is obtained from each request checked with GDTS from backend. If it is the same a "304 – Not Modified" response is sent back, asking the client to use the local cache copy. If it is not exactly same, a the request is passed through and rendered dynamically. This guarantees all content will be served fresh after every deployment. Requests in between deployments may be served from client-side cache, because they will have same ETag.
However, having just ETag will still generate "GET " request and "304 – Not Modified" response (ping) conversation for each request, even if it has the latest version in the cache. Meaning, there will still be network round trips for each page. So a Cache-Control header is also used. This header, Cache-Control: public, max-age:3600. caches the content at the BrowserAgent (client) for an hour. So the change pings will not called at least for 1 hr.
This way, it is guaranteed that data can be stale at most 1 hr. (can also be considered downside). However, according to the site's analytics, max length of time spent by a user on the site is 14 mins. 1 hr. covers the most dedicated visitor :)