Logpipe: read an unfamiliar log pipeline, fix three reports, add a feature, then make ingest fast

Context

You have just joined the observability team, and logpipe/ is yours now. Nobody who wrote it is around. Read README.md and the code first: the reports below assume you know which file does what.

Work through the three phases. Every check is graded on its own, so each fix counts even if you do not finish.

Phase 1: debug

CMP-1: WARN lines are dropped

Report: The search team's warnings never reach the dashboard. ingest("50 WARN search /query 210 r-1") returns False.

• Lines at every level the README lists are taken.

CMP-2: p95 reports the slowest request

Report: With exactly 20 requests to /query, taking 10, 20, ... 200 ms, p95("/query") says 200. By the README's nearest-rank rule it is 190.

• p95 follows the nearest-rank rule for any number of requests.

CMP-3: Error rates stay high after an incident

Report: Checkout logged errors at 10:00, then went quiet. At 10:05 error_rate("checkout", now) still counts them.

• error_rate(service, now) counts only the service's lines in (now - 60, now], even when the service has sent nothing since.

Phase 2: extend

CMP-4: The slowest endpoints

Add slowest_endpoints(k) to LogPipeline: the k endpoints with the highest p95, as [endpoint, p95], highest first. Equal p95s go by endpoint, A to Z. With fewer than k endpoints, return them all. Build it on the pipeline's existing latency stats.

• slowest_endpoints(k) returns the endpoints as described.

Phase 3: optimise

CMP-5: Ingest slows down as the day goes on

Report: Replaying an hour of logs (about 20,000 lines) through ingest_many takes many seconds, and each line is slower than the one before. Find out why and fix it without changing what any method returns.

• 20,000 lines are taken within 1 second, the time limit of the speed checks.

Notes

• Run pytest tests/visible. Three tests fail on the current code; not every report has a failing test.
• In the reported round candidates may use an AI assistant. Whatever you use, be ready to explain every change you make.
• Practice task in the format reported for Google's code comprehension interview round, piloted in 2026. Original StackYard code, not an actual Google question; not affiliated with Google.