Practice: A Cache Policy for a News Site

Headlines change hourly, articles almost never, and the homepage every minute.

Invalidation, eviction and the stampede are practised together on one news site here. Write down a policy for each kind of value before opening the answer.

The brief. A news site caches three kinds of value in a shared Redis cache with 8 GB of memory.

Articles: hundreds of thousands, written once and corrected rarely; a correction must show at once.
Section headlines: a list per section, updated about hourly by editors; a few minutes late is acceptable.
The homepage: built every minute from the headlines, read 5,000 times a second.

Which invalidation strategy suits each?
Which articles should the cache hold, and what happens when it is full?
What could go wrong when the homepage entry expires, and how do you prevent it?

Check your policy

Articles: delete on write, so a correction shows on the next read, with a long TTL as a backstop. Headlines: a TTL of a few minutes is enough, since a short delay is allowed. The homepage: a TTL of about a minute, matching how often it is rebuilt.
Hundreds of thousands of articles will not all fit, and most are rarely read after their first days. LRU eviction with a memory limit keeps the articles being read now and drops the old ones.
A stampede. At 5,000 reads a second, every request in the rebuild window would rebuild the homepage at once. Let one request rebuild while the others wait or are served the previous version.

← → move through sections and cards