Skip to content
Tech Interview Prep
Technical interview guide

Bioinformatics Pipelines & Reproducibility

How multi-tool genomics analyses are made repeatable and portable: workflow managers, containerization, and the reproducibility practices the field has had to learn the hard way.

Read
50 min
Practice MCQs
25
Interview Q&A
25
Edition
v3
Editorial status
Reviewed

Scope: Nextflow 25.x; Snakemake 9.26; CWL 1.2.1; WDL 1.2; OCI image specification; current nf-core, RO-Crate 1.1, FAIR and GA4GH WES guidance reviewed 2026-09-04.

Curated: · Written: · Reviewed:

25 questions

MCQ-9A service repo contains `requirements.in` (`flask`, `requests`, no version bounds) and `requirements.lock` (fully pinned, generated by `pip-compile`). CI runs `pip install -r requirements.lock`; two builds from the same commit, three days apart, both install `urllib3==2.1.0` and `werkzeug==3.0.1`. A developer runs `pip install -r requirements.in` on their laptop, gets `urllib3==2.2.1` and `werkzeug==3.0.3`, and the app then fails at import with an error raised inside a transitive dependency. What explains the difference?
MCQ-7A pipeline ingests signed model artifacts from an internal registry. Each artifact ships with a SHA-256 digest recorded in an immutable manifest at publish time. After a disk-level bit flip during transfer, the artifact's bytes change but its manifest entry does not. A verification step recomputes the digest and compares it to the manifest. What does this verification establish about the artifact?
MCQ-10A team's Nextflow pipeline hard-codes `executor 'slurm'`, `queue 'gpu-long'` and an absolute `s3://prod-scratch/` staging path inside each process body. The same pipeline must now run on the company's Kubernetes cluster and on a laptop for a demo, with no edits to process bodies. Which of these should move out of the process definitions?
MCQ-24A rare-disease consortium holds 8,200 exomes behind a Data Access Committee. Its landing page carries a DOI and a DataCite record whose machine-readable fields are the title, the DOI and a link to a PDF; the phenotype summaries, per-sample consent scope and DAC application procedure exist only inside that PDF. A federated search filtered on disease area and consent scope returns nothing for this dataset. Which change closes the gap the filter exposes?
MCQ-13A CI golden test hashes a JSON report the service emits and compares it to a stored hash. On unchanged code it fails about 1 run in 5. The hash is computed over the report after `generatedAt` has been removed. Diffing a failing run against the golden file shows the same tag keys and values, but the keys appear in a different order each run; the report is built by ranging over a Go `map[string]string` and marshalling it. Which change stops the flake without weakening the test?
MCQ-14A team records the seed for a training run, then reruns the identical script on a different GPU node with a newer cuDNN build; the loss curve diverges. Which of these did recording the seed not guarantee?
MCQ-1A team runs a Snakemake pipeline on a shared cluster. Rule `align` declares `input: "ref.fa", "reads/{sample}.fq"` and `output: "bam/{sample}.bam"`; rule `call` declares `input: "bam/{sample}.bam"` and `output: "vcf/{sample}.vcf"`. A new engineer updates `ref.fa` by restoring it from a backup tarball with `tar -x`, which writes the new contents but restores the archived modification time: older than the existing BAMs. They rerun `snakemake -j 8`. Snakemake reports nothing to be done and the old VCFs remain. What is the mechanism behind this outcome?
MCQ-15A CI build cache keys object files by SHA-256 of the source bytes plus the exact command line. The command is `c++ -O2 -c foo.cpp`; `c++` is a symlink at `/usr/bin/c++`. The image is updated from GCC 11.4 to GCC 13.2, but the symlink path and command line are unchanged. The cache still hits, and CI links objects compiled by GCC 11.4; a miscompilation fixed in GCC 13.2 still reproduces. Which change to the cache key prevents this reuse?
MCQ-8A Deployment in git reads image: registry.internal/pay-api:2.7.1, and CI pushes that same tag on every merge to main. The 2025-03-04 release was load-tested at 1200 rps with p99 of 38 ms, and its digest sha256:aa11c7 is written in the release ticket. On 2025-03-21 an engineer re-applies that same March 4 manifest to roll back a config change. The pods start, but p99 is 90 ms, startup logs print 2.7.1+build.8842 where the March pods printed 2.7.1+build.8711, and every restarted pod reports imageID ...@sha256:bb220f. Nodes, cluster version and resource limits are unchanged between the two runs. Which explanation fits the evidence?
MCQ-21A team's CI runs ESLint, `tsc --noEmit` and `terraform validate` on every PR; all 58 PRs in six weeks passed. Of the 58 deploys that followed, 41 failed within 90 seconds of container start: 33 with `Cannot find module 'sharp'` and 8 with `ImagePullBackOff` after `latest` was overwritten by a later build. Which change makes a green pipeline run mean the exact artifact that ships has started successfully?
MCQ-3A nightly job writes one Parquet file to S3, and the orchestrator marks the task successful as soon as the file exists and is larger than zero bytes. Last night it reported success with a 12 MB file holding 128,000 rows. The trailing 30 runs averaged 380 MB and 4.2M rows; the file parses cleanly and its columns match the registered schema. Downstream, the day's revenue reads zero. Which change to the task contract would have caught this before the task was marked complete?
MCQ-19A nightly variant-calling pipeline runs 400 samples. Each task reads a FASTQ, writes a VCF, and exits 0. Monitoring records exit status, wall time, and peak RSS. Last night all 400 tasks exited 0, wall time was 18–24 minutes for every task, and peak RSS was 3.1–3.4 GB for every task. Downstream, 12 of the 400 VCFs hold zero variant records while the other 388 hold 40,000–52,000. The 12 inputs had been truncated during transfer, and the transfer tool exited 0 too. Which monitoring addition would have flagged exactly those 12 tasks when they finished?
MCQ-4A sequencing core validates each sample sheet only after demultiplexing. Last week a sheet listed `ACGTACGT` as the `index_sequence` for two different `sample_id` values; bcl2fastq wrote both samples into one FASTQ pair, and the collision surfaced at alignment. By then the 9-hour NovaSeq run and the 40-minute demultiplex had already been spent. The demultiplexer is a vendor binary the core cannot reconfigure. The core wants this class of error caught before the run is queued. Which change does that?
MCQ-20A Snakemake RNA-seq pipeline has six rules. CI runs it with stub scripts that create empty files at each rule's declared output path, then asserts every output exists; the suite is green. The first production run on 40 samples fails at the counting rule with "no reads in BAM": the aligner's config default writes CRAM, and the metadata parser silently dropped the sample sheet's strandedness column. CI must finish in under 10 minutes. Which fixture change would have caught both failures before production?
MCQ-5A batch tool builds its effective configuration from three layers, each overriding the one before: `defaults.yaml`, then the profile named by `--profile` on the command line, then environment variables prefixed `APP_`. `defaults.yaml` sets `retries: 1`; `profiles/prod.yaml` sets `retries: 5`. A production run invoked as `tool --profile prod` performed 2 retries, and the run log records only that command line. Which explanation fits, and what should the team change?
MCQ-11A Nextflow 24.04 `align` process declares `memory '8 GB'` and `cpus 2`. Profiling 40 pilot samples with 2–25 GB inputs fits peak RSS ≈ 3 GB + 1.2 × input size. In production the scheduler OOM-kills (exit 137) about 70% of tasks. What should the memory declaration be?
MCQ-16A job runner processes 2,000 orders. Each job calls POST /charge with Idempotency-Key: order_id. The API documents: 200 = charged; 400 = invalid request; 409 = a request with the same idempotency key is still in progress, retry with the same key; 503 = transient, retry with backoff. The runner's policy: max 3 retries, backoff 1s, 2s, 4s; total job deadline 10s. Job 812 fails on attempt 1: POST /charge returns 409 after 0.5s. What should the runner do?
MCQ-2`load_orders` runs once a day. It reads a paginated REST endpoint, 500 records per page, ordered by `updated_at`, and writes every record into `orders_stage` with a plain `INSERT INTO`, partitioned by run date. The orchestrator guarantees at-least-once delivery: a lost ack can hand the same task to a second worker while the first is still running, and a task whose worker dies is re-run from page 1 on another worker. On 2024-03-11 a clean run writes 8 pages. The worker died after committing pages 1-3, the task was re-run from page 1 elsewhere and finished all 8 pages, and `orders_stage` now holds 5,500 rows for that date. The source is written to while the job runs, so records move between pages from one attempt to the next. Which change makes a re-run end with the 4,000 rows a clean run produces?
MCQ-12A nightly scatter-gather job fans out over 512 shards keyed by customer_id, each shard writing one Parquet file under date=2024-06-01/. Before publishing, the gather step builds a manifest of shard keys. Last night 511 shards finished on the first attempt; shard 037 timed out, and its retry succeeded, but the timed-out attempt had already flushed a partial file. The manifest now lists 513 files covering 512 distinct keys, with shard 037 present twice. What should the gather step do?
MCQ-22A clinical variant-calling pipeline caches per-sample VCFs keyed by sample ID and must keep every reported result reproducible for a 2-year retention window. The lab replaces the reference FASTA at the same path and filename, adding a patch contig and reordering contigs, then re-runs 40 samples. The cache still holds entries for 37 of them, so those 37 return their stored VCFs unchanged; the 3 with no entry are recomputed against the new reference. Which change fixes the pipeline without breaking reproducibility?
MCQ-25A clinical lab runs a 480-sample variant-calling pipeline as 96 shards on a shared cluster with resume-from-cache enabled. Last month a run exited 0 with every task green, but the cohort report was wrong: 12 samples had been served from cache entries keyed only on sample ID, so a corrected input file at the same path still hit the old cache, and the only output check was a final VCF row count. Which release gate would have caught this before the report shipped?
MCQ-17A nightly pipeline runs a discovery task that lists the day's new S3 prefixes, then fans out one training task per prefix. The run is checkpointed after discovery, and the checkpoint stores the prefix list alongside the task graph. A week later the team replays that checkpoint to reproduce Tuesday's model. The replay succeeds, but the resulting model differs from Tuesday's. Discovery is re-run against today's bucket, which now holds 40 new prefixes. What is the most likely cause?
MCQ-18A nightly pipeline writes one provenance row per task attempt and prints one stdout line per task. Tuesday's rows for the training run: | task | input digest | output digest | config hash | attempt | | --- | --- | --- | --- | --- | | ingest | raw-9f2 | ds-4a1 | c7 | 1 | | train | ds-4a1 | model-11b | c7 | 1 | | train | ds-4a1 | model-3e8 | c7 | 2 | | eval | model-11b | metrics-77 | c7 | 1 | The team promoted model-3e8. Wednesday's eval shows a 9-point AUC drop. Tuesday's stdout shows every task 'finished successfully' with no warnings. What does the provenance table establish that the stdout cannot?
MCQ-6A rerun and its archived baseline were aligned with the same aligner build and the same command line, and both BAMs record AS:GRCh38 with contigs named chr1, chr2, and so on. Depth agrees everywhere except 12 kb of a duplicated chr1 segment: the rerun places those reads on the ALT contig for chr1, the baseline places them on chr1 itself, and one variant call flips between the two. Which explanation fits?
MCQ-23A rare-disease study's sample S-1042 fails secondary analysis at 03:14. The pipeline writes task logs to a CloudWatch group that a nightly job exports to a public status dashboard; each line contains run_id, sample_id, workflow, and exit code. The dashboard must stay public for uptime, and engineers must debug failed runs by run_id. Which change best prevents the dashboard from linking S-1042 to that participant's restricted variant calls while keeping failed-run debugging?