Translation Queue

The Translation Queue is a real-time view of every translation job Sublarr is working on right now, plus the most recent finished and failed ones. Jobs stream their progress over WebSocket so the cue counter ticks up cue-by-cue without polling. From the queue you can cancel a runaway job, retry a failure, or jump into the resulting subtitle in the editor.

Where it lives

Open it from:

Surface	Notes
Settings → Translation → Queue	Dedicated page; the most complete view.
Activity → Translations tab	Same data with a smaller header for quick checks.
Dashboard → Activity Feed	Translation entries link straight to the queue page.

Layout

The page stacks three sections:

Section	Content
Active jobs	Jobs in running or queued state. Live progress bars.
Recent completions	Last 50 finished jobs with cost, duration, and link to the result.
Failures	Jobs that errored, with truncated error message and retry button.

Job card anatomy

Each job in Active shows:

Field	Effect
Series / Movie	What’s being translated. Click to open the source media in Library.
Source → Target	Language pair (en → de).
Backend	Which translation backend is doing the work.
Progress bar	Percentage based on cues completed / total cues.
Throughput	Cues/second over the last 60 s rolling window.
ETA	Estimated remaining time at current throughput.
Cost (running)	Tokens or characters consumed so far × backend’s per-unit price. Hidden for free backends.
Cancel	Cooperative cancellation — finishes the in-flight cue, rolls back partial output, marks cancelled.

Tip

ETA is a moving estimate. If throughput drops mid-job (network glitch, backend rate-limit), ETA jumps; if it speeds up (cache hit run), ETA shrinks. Use the trend, not the absolute number, to decide whether to cancel.

Recent completions

Each row links to the actual subtitle the job produced. Sortable by:

Column	Default sort
Finished at	Most-recent first (default).
Duration	Longest first — surfaces slow jobs.
Cost	Highest first — surfaces expensive jobs.
Cues	Largest first — proxy for content size.

Clicking a row opens the Subtitle Editor on the produced file, with the source side-by-side for comparison.

Failures

Failed jobs surface inline so you don’t have to dig through logs:

Column	Effect
Series + episode	Same as active.
Error	One-line summary (e.g. “Backend rate-limited”, “Cue exceeds context”).
Backend	Which one failed; useful for “switch and retry” decisions.
Logs	Opens the relevant section of Logs filtered to the job ID.
Retry	Re-queues the job with the same parameters.
Retry on different backend	Picks the next backend in the fallback chain and queues.

Cancellation

Cancel is cooperative, not forceful:

1. The cancel signal is recorded in the job state.
2. The current cue’s translation finishes (in-flight HTTP request completes).
3. Partial cues already written are dropped — no half-translated file lands.
4. Job exits with state cancelled.

Total time from click to terminal state is bounded by your backend’s per-request timeout (default 90 s).

Concurrency

The queue respects Settings → Translation → Workers — usually 2 concurrent jobs. Additional jobs queue with state queued and start when a slot frees. The queued count appears in the header so you can see backlog at a glance.

Cost telemetry

Every completed job writes a row to translation_events with cost, latency, token / character counts, and backend ID. The aggregated view lives in Cost & Memory; the queue shows the per-job number for situational awareness.

Filtering

A toolbar above the lists supports:

Filter	Effect
Status	Show only running, queued, completed, failed, cancelled.
Backend	Only jobs from one backend — useful when comparing quality between backends.
Series	Only jobs for one show.
Date range	Limit the recent-completions list.

Persistence

The queue is durable across restarts. If Sublarr is stopped mid-job, the job re-enters queued state on next boot and resumes when a worker picks it up — translation memory ensures no cue is paid for twice.

Caution

“No cue paid for twice” assumes the same backend + same source language + same target language. If you change the backend mid-job (cancel + retry on another), the new backend pays the full cost — even for cues the previous backend had already translated.

When something looks stuck

Symptom	Where to look
Job is queued for hours	Workers maxed; raise Concurrent translations in Settings → Translation.
Throughput is zero on running job	Backend not responding — check Settings → Translation → Backend health + Logs.
Job shows failed with empty error	The error message was longer than the truncation limit; click Logs for the full trace.
All jobs failing on one backend	Auth issue (rate limited / invalid key) — check API Keys → Test for that backend.