Part 9: Kafka Benchmark Review Quiz
Distinguish reported measurements, causal claims and reproduction checks.
1. What is the storage model for three brokers with RF3, and how should F1's 134.74 MiB/s be interpreted?
Show answer
Each broker holds one copy of every partition. Encoded-log and client-payload rates differ; F1 is a short single-run client report. Reconcile cache, writeback, windows and units before claiming a validated sustained 130–135 MiB/s ceiling.
2. Do low IOPS and high byte rates prove that volume throughput is the only bottleneck?
Show answer
No. They are consistent with sequential I/O/throughput pressure, but instance EBS baseline/burst, cache, clients and sampling uncertainty also matter. Do not equate a 135 peak sample with a provisioned 125 MiB/s value. Test the hypothesis with controlled changes.
3. With B1/B2 p50 at 3 ms and p99 at 126/17 ms, can you conclude that acks affects only the tail?
Show answer
No. The p99 ratio is about 7.412, but means also differ at 5.27/2.58 ms. Equal sampled integer-ms medians do not prove equal underlying latency, and one pair is not universal. An acks=0 callback is not a broker durability acknowledgement.
4. Why is F4/F1's 337.81/134.74=2.507 ratio not a pure replication-cost measurement?
Show answer
Both RF and acks change, from RF3/acks=all to RF1/acks=1, and omitted idempotence defaults differ too. Duration, per-broker data, cache and client constraints also differ. Preserve the observed ratio without presenting it as an isolated RF effect.
5. How should F6 be described when batch size, linger and record count all change?
Show answer
148.38 MiB/s is 10.12% above 134.74, with a lower reported CPU range. The 39.19% CPU-midpoint difference is arithmetic, not a batch-only effect. Without repeated variance measurements, the throughput difference cannot be classified as noise.
6. Can the padded JSON's compression ratios and codec ordering be transferred unchanged to other logs?
Show answer
No. About 63.2% x padding makes the distribution unusual; neither ratios nor ordering are universal bounds/rankings. Regenerating the AWK with another implementation can change hashes and compression. Retain the original payload, image and tool versions.
7. Which confounders matter for E4's slowdown and the EC2 network counters?
Show answer
The producer falls 44.49%, from 103.36 to 57.37 MiB/s, but shares a 1.9-CPU client/NIC with the consumer. Low broker CPU does not exclude client contention. EC2 rx/tx credits are separate: F1 rx136 MiB/s≈1.141 Gbps and tx108≈0.906 Gbps are each below 1.25 Gbps. Do not add them to claim bursting.
8. How should a new reproduction validate units, steady state and successful delivery?
Show answer
30M×1024 B is 28.610 GiB, and no universal ten-GiB/one-minute steady-state threshold exists. Record warmup, long windows, drain, repeats, common denominators and actual device/timestamps. Check modern options, fresh populated topics, final successful producer count/stderr and consumed contents. Callback failures may not be captured by exit status alone; a cost estimate is not a bill.