Day 25 results: duplicates, dedupe, and the honest middle

Measured on: (machine, OS, Python version)

Output

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

One sentence per prediction

P1 (at-least-once duplicate charges):
P2 (idempotent duplicate charges):
P3 (at-most-once lost charges):
P4 (at-least-once total deliveries):

Why the duplicates happened

Same crashes, process-then-ack. How many duplicate charges, and what was the consumer doing on every redelivery:

What the idempotency key changed

The broker redelivered exactly as often as in Part 1. So why did the ledger come out exactly right this time:

The dilemma in Part 3

Ack-first lost messages, process-first duplicated them. In one line, why can you not just do both steps atomically and get neither problem:

The number that matters

Without a key, __ duplicate charges (ledger too high). With a key, __ duplicates (exactly right), same crashes. Why that is the real meaning of “exactly-once”: