Parallelism and Indexed Work

Two separate numbers control a Job's scale, and it is worth being precise about which is which: completions is the total number of successful work units the Job needs before it is considered done, while parallelism is how many Pods are allowed to run at the same time working toward that total. A specialized variant, an Indexed Job, additionally gives each Pod a stable, predictable completion index (0, 1, 2, ...) that the application can read - useful for cleanly splitting a large dataset into fixed shards where each worker knows exactly which slice is "its" work. Not every batch workload fits that clean, known-count model, though - many real work-queue patterns instead have an unknown, growing number of items and coordinate through an external queue rather than a fixed completions number. Whichever pattern you use, remember that parallelism is a knob you control, but the systems your Pods talk to - a database's connection pool, an external API's rate limit - are not infinitely elastic just because Kubernetes can create more Pods; concurrency has to respect the narrowest downstream limit in the chain.

Analogy: Kubernetes is a thermostat, not a remote control. You declare the temperature; independent controllers measure reality and keep acting. Debugging means finding which sensor, rule or actuator prevents convergence.

Read the object as evidence

kubectl get RESOURCE NAME -n NAMESPACE -o yaml
kubectl describe RESOURCE NAME -n NAMESPACE
kubectl get events -n NAMESPACE --sort-by=.metadata.creationTimestamp
kubectl explain RESOURCE.spec

Do not memorize these as a ritual. The first command exposes desired and observed state, the second connects conditions and events, the third supplies a timeline, and the fourth checks the server's schema. Compare what the controller was asked to do with what it reports doing. Then test the narrowest hypothesis.

Scenario: A database-migration Job's parallelism is raised from 10 to 100 Pods in an attempt to finish faster, but the shared connection pooler (PgBouncer) in front of the database can only sustain a fixed number of concurrent connections. The extra Pods spend most of their time waiting for a connection slot rather than doing useful work, total throughput actually falls, and the migration takes longer than it did at 10 parallel Pods. More Pods is not automatically more throughput - it only helps up to whatever the narrowest real downstream bottleneck can absorb.

Bound parallel work

completions is required successful units; parallelism is simultaneous Pods. Indexed Jobs give each Pod a stable completion index for deterministic shards, but the application must map indexes to non-overlapping work. Account for database connections, API rate limits and cluster capacity before raising parallelism. Watch succeeded, failed and active counts rather than only Pod phase.

Scenario: Parallelism rises from ten to one hundred and total throughput falls because every worker contends on one database lock. Kubernetes delivered concurrency correctly; the dependency was the bottleneck.
Tip: Start below the downstream concurrency budget, measure queue latency and error rate, then increase gradually. Pod count is a control input, not a throughput guarantee.

Production reasoning

Ask four questions: Who owns this object? What dependency must become ready next? Which controller reports the blocking condition? What evidence would disprove my current theory? This prevents symptom-driven changes. Record the context, namespace, object generation, image digest and recent rollout before mutation; a recreated Pod may erase the evidence you needed.

Warning: Running is not the same as ready, healthy, durable or correct. Kubernetes status is layered. Confirm the application-level outcome as well as the object state.
Goal: Put this model into practice in the Kubernetes lab init-pgbouncer-postgres. Open /labs/kubernetes, choose init-pgbouncer-postgres, predict the failure path before changing anything, then use the simulator's check command to validate the finished state.

Deliberate practice

Before the lab, write the expected object relationship and the first three commands you will run. Afterward, explain why the fix converged and name one tempting change that would only mask the symptom. Repeat using an explicit namespace and a structured output format. This prediction-observation-explanation loop is what turns command familiarity into production judgment.

45-minute investigation

  1. Map (5 min): draw the owner-to-child chain and mark every namespace, selector, identity and dependency involved.
  2. Predict (5 min): write one expected status condition, one likely event and one log or metric signal before opening the lab.
  3. Observe (10 min): collect YAML, describe output and ordered events. Do not mutate state. Record which observation disproves your first theory.
  4. Repair (15 min): make the smallest declarative correction, watch the responsible controller converge, and verify the user-facing path rather than stopping at Running.
  5. Stress (10 min): change one relevant constraint - replica count, label, readiness, resource value or placement rule - predict the outcome, observe it, then restore the known-good declaration.
Tip: Keep a short incident note with symptom, evidence, hypothesis, change, result. Across four sections this produces a reusable runbook instead of a pile of remembered commands.