Added short-term caching of the first column in the channel manifest

[billkalter]

Github Issue

None

What Are We Doing Here?

An inefficiency was uncovered when the head of the "manifest" table for queues and databus subscriptions have accrued many tombstones which have not yet been compacted away. Since every poll re-reads the manifest from the beginning the entire poll is slowed down as Cassandra scans over the tombstoned records. This PR briefly caches the oldest know slab ID in the manifest plus a 1 minute buffer. Future queries can be run starting at this manifest to bypass any tombstones from older, fully read and deleted slabs.

How to Test and Verify

There is no test specifically for this condition. The most important test is regression.

Risk

This is a fairly low-risk update. Even though it is at the heart of databus and queue channels the caching should serve as an optimization without risking that any manifest data goes completely unread.

Level

Medium

Required Testing

Regression

Code Review Checklist

• Tests are included. If not, make sure you leave us a line or two for the reason.

• Pulled down the PR and performed verification of at least being able to
  build and run.

• Well documented, including updates to any necessary markdown files. When
  we inevitably come back to this code it will only take hours to figure out, not
  days.

• Consistent/Clear/Thoughtful? We are better with this code. We also aren't
  a victim of rampaging consistency, and should be using this course of action.
  We don't have coding standards out yet for this project, so please make sure to address any feedback regarding STYLE so the codebase remains consistent.

• PR has a valid summary, and a good description.

[sujithvaddi]

@billkalter should it be put() here instead of get() :?

[billkalter]

Agreed, this is very confusing and I tried to put a comment to clarify. With the newer Java interfaces like ConcurrentMap you can use computeIfAbsent() or putIfAbsent(). The Guava cache interface doesn't have a similar method, but if you do a get() this way it does the same thing: caches the new value only if there is no current un-expired version in the cache. From their docs:

This method provides a simple substitute for the conventional "if cached, return; otherwise create, cache and return" pattern.

[sujithvaddi]

@billkalter looks good.