Bash

Use flock to prevent overlapping Bash jobs without stale PID files

The Tryssh team ·

For flock Bash single instance, begin by proving which system and state you are about to affect. Flock asks the kernel to hold an advisory lock on a file descriptor or path for the process lifetime, avoiding manual stale-PID cleanup. The safest useful answer is therefore an evidence sequence: identify the target, capture its effective state, make one bounded change, and verify through a path independent of the command that accepted the change.

TL;DR: One invocation holds the lock, a second exits with a documented status, and the lock releases after process death. Start with pgrep -af 'example-job.sh', stop if the evidence instead shows that different paths or namespaces create separate locks, a child retains the descriptor, or callers ignore lock-failure status, and keep this recovery action ready: restore the prior scheduler command, terminate only the verified candidate process, and define one shared lock path for every writer.

Assumed audience: operators writing unattended Bash scripts that must fail clearly, preserve secrets, and return meaningful status to a scheduler. This field guide assumes a controlled maintenance window and working recovery access.

Use flock to prevent overlapping Bash jobs without stale PID files identity, evidence, action, and proof map

Direct answer and success condition

The success condition for flock Bash single instance is not a zero exit code. It is that one invocation holds the lock, a second exits with a documented status, and the lock releases after process death. A command may return successfully while a controller is still converging, a client is using cached state, a remote dependency is unavailable, or the wrong account, namespace, region, host, database, repository, or process accepted the request.

Use the task boundary—parsing, expansion, quoting, pipeline status, subprocess lifetime, signals, cleanup, logging, and scheduler-visible exit status—to decide what the result proves. The candidate action belongs only at the first layer whose observed state contradicts the desired outcome. If the baseline cannot locate that contradiction, do not make the action broader.

Working model

For Bash automation, use this operational model: Bash parses shell syntax, performs several expansions, executes builtins or child processes, and derives compound-command status according to context and enabled options. This model matters for flock Bash single instance because flock asks the kernel to hold an advisory lock on a file descriptor or path for the process lifetime, avoiding manual stale-PID cleanup. It also separates four forms of evidence that are often collapsed:

  1. Declared state: files, manifests, playbooks, policies, arguments, or API requests say what should happen.
  2. Effective state: the running tool or service reports what it actually loaded and selected.
  3. Resource state: processes, objects, data, network paths, and queues reflect the change.
  4. Outcome state: the original user, automation, or recovery workflow succeeds from the relevant vantage point.

Declared state without effective state is only intent. Effective state without outcome state is only partial convergence. Keep these labels in the change record so another operator can tell what was measured.

Preflight: identity, scope, and recovery

Test with representative empty, whitespace, failure, signal, and timeout cases; use a disposable target and never enable tracing around secrets.

Record the current UTC time, operator identity, tool version, exact target identifiers, last successful execution, recent related changes, and the owner of the workload. Save the effective configuration or object state using its supported read-only interface. Do not store credentials, complete environment dumps, private customer data, or unrestricted topology in the ticket.

The recovery path for this guide is explicit: restore the prior scheduler command, terminate only the verified candidate process, and define one shared lock path for every writer. Rehearse the targeting syntax and confirm that recovery does not depend on the same account, network path, key, state file, database, repository, or process being changed.

Stop before changing state when any of these statements is true:

  • The target can be selected by a default or ambiguous alias.
  • The current configuration or binding cannot be reconstructed.
  • The only privileged or remote session would be put at risk.
  • The command affects an unbounded host, key, object, snapshot, branch, or resource set.
  • The expected output cannot be distinguished from stale, cached, or partial state.

Capture the baseline

Run the following commands individually after replacing example identifiers deliberately:

pgrep -af 'example-job.sh'
lslocks | rg 'example-job.lock' || true
systemctl list-timers --all | rg example

For flock Bash single instance, classify the output before proposing a fix:

  • Expected evidence: one invocation holds the lock, a second exits with a documented status, and the lock releases after process death.
  • Abnormal evidence: different paths or namespaces create separate locks, a child retains the descriptor, or callers ignore lock-failure status.
  • Inconclusive evidence: no output, permission errors, incomplete history, disabled instrumentation, a different version, or a different control plane can all hide the relevant state. Confirm those assumptions rather than translating absence into health.

Preserve timestamps and exit statuses for decisive observations. Prefer machine-readable output when it can be filtered without collecting secrets. Compare a healthy peer only by equivalent effective fields; copying its entire configuration can introduce a second problem.

Controlled action

This is state-changing example syntax. It is intentionally presented after the baseline and must not be pasted with example targets:

flock -n /run/lock/example-job.lock /usr/local/bin/example-job.sh

The action is justified only if the baseline predicts its effect at the named boundary. For flock Bash single instance, the principal risk is that locking only one entrypoint does not stop another tool from mutating the same resource. Review the exact expansion of variables, globs, resource addresses, inventory patterns, database identities, repository locations, and cloud regions before approval.

Prefer a canary, dry run, saved plan, isolated restore, configuration validator, transaction, immutable artifact, or runtime drain when the platform supplies one. Record the command, approver, UTC time, and expected convergence interval. Do not stack unrelated cleanup, restart, permission, and configuration actions into the same observation window.

Independent verification

Repeat the state inspection and then exercise the original path:

lslocks | rg 'example-job.lock' || true
journalctl -u example-job.service --since today --no-pager
pgrep -af 'example-job.sh'

Verification for flock Bash single instance must answer five questions:

  1. Did the intended identity accept the operation?
  2. Did effective state converge to the reviewed value?
  3. Did the underlying resource or data path change as predicted?
  4. Did the real consumer succeed from an independent vantage point?
  5. Did adjacent safety signals—capacity, latency, errors, replication, audit, or persistence—remain healthy?

If lslocks | rg 'example-job.lock' || true passes but the consumer still fails, the local boundary may be repaired while another layer remains broken. Keep the new evidence, stop making the change broader, and move to the next falsifiable boundary.

Decision branches

The baseline contradicts the guide

If different paths or namespaces create separate locks, a child retains the descriptor, or callers ignore lock-failure status, the candidate action no longer follows from the evidence. Re-establish identity and scope, shorten the observation window, and formulate a mechanism that the next read-only command can disprove.

The command succeeds but nothing converges

A successful flock -n /run/lock/example-job.lock /usr/local/bin/example-job.sh proves only that one interface accepted the request. It may not prove persistence, controller completion, process reload, data compatibility, replication, traffic admission, or client refresh. Inspect those transitions in order.

The change makes the outcome worse

Execute the prepared recovery: restore the prior scheduler command, terminate only the verified candidate process, and define one shared lock path for every writer. Preserve the failed candidate, event times, and relevant logs. Avoid repeated restarts, broad resets, garbage collection, pruning, history rewriting, or cleanup that can erase the evidence needed to explain the failure.

The result is mixed

Mixed results normally mean scope differs across hosts, workers, replicas, zones, clients, branches, or repositories. Partition the evidence by identity instead of averaging it. Hold further rollout until each partition has an explicit disposition.

Review checklist

  • Confirm the operator, account, region, namespace, host, resource, repository, database, or branch.
  • Print the installed tool version and resolve the effective configuration.
  • Capture the baseline and name one observation that would disprove the proposed mechanism.
  • Mark flock -n /run/lock/example-job.lock /usr/local/bin/example-job.sh as state-changing in review.
  • Keep recovery access independent and test the exact rollback target.
  • Change one layer and wait for its documented convergence boundary.
  • Repeat the original user or automation path, not only the control-plane query.
  • Watch error rate, latency, capacity, data durability, audit, and persistence after the change.
  • Remove temporary credentials, traces, restored data, debug settings, and candidate resources under policy.
  • Update this field guide when observed behavior differs from the source-reviewed model.

Investigate it in Tryssh

Tryssh keeps the target, read-only evidence, approval boundary, command output, and recovery decision in one host conversation.

Tryssh can help preserve this evidence trail and place a state-changing command behind human approval. It cannot decide the correct production target, authorize a cloud or database change, guarantee backup completeness, or replace independent recovery access.

Evidence and review status

This flock Bash single instance field guide was source-reviewed on 2026-07-29 against current upstream or first-party documentation. Commands use example identifiers and were not executed against every distribution, service version, provider, database topology, repository backend, network, or workload. Provider-managed services may expose a different control plane or restrict local commands.

The article makes no ranking guarantee and does not treat documentation review as reproduction. Validate installed versions, permissions, feature support, recovery behavior, and billing or data-retention consequences in your environment.

Limitations and trade-offs

Locking only one entrypoint does not stop another tool from mutating the same resource. A narrow safe action may take longer than a broad reset, and a strong verification plan may require temporary capacity or an isolated restore target. Those costs are part of reliable operations, not optional ceremony.

Do not use a search result as authority to change production. The live system, reviewed policy, upstream versioned documentation, and accountable operator remain the sources of truth.

Continue the cluster

Next, read timeout remote command or Bash read lines safely. For broader context, use the Bash automation foundation guide and the SSH hardening checklist.

Sources and further reading