Day 12 results: replication lag and the stale read

Measured on: (machine, OS, Python version)

Output

Paste part 1, part 2, part 3 and the scoreboard here.

One sentence per prediction

P1 (stale percent at the given lag):
P2 (stale window in ms):
P3 (one synchronous write in ms):
P4 (stale percent after the fix):

The stale read, in my own words

At 40 ms of lag, what fraction of my immediate follower reads were stale, and how long did the follower keep serving the old value:

Which fix I would reach for

Read your writes (route recent reads to the leader) or wait for the follower to catch up, and why, for a feed like this:

The number that matters

Stale percent at the given lag was __ and after the read-your-writes fix it was __. Synchronous writes cost __ ms versus __ for async. The tradeoff in one line: