v5.1.0 (#210)

• Fixes

  • A comic with no issue number no longer mis-tags itself with the first
    issue of a series matched earlier in the same run.
  • Metron tagging no longer sends requests the server is going to refuse.
    Every request now passes through a process-wide rate gate that models
    Metron's own sliding window, so a -j N batch stays inside the per-minute
    limit instead of discovering it by being rejected. This matters more than
    it sounds: Metron counts a rejected request against the daily quota
    exactly like a successful one, so the old behavior spent budget on
    nothing. Reported as
    #207.
  • A rate-limit rejection now releases one worker when one slot frees, rather
    than every worker at once. Retry-After says when a single request
    becomes possible, not when the window refills; waiting it out and then
    sending the whole queue earned one success and a fresh burst of
    rejections.
  • online.tuning.retry_budget is wired up. It was parsed, defaulted and
    then never passed to the retry decorator, so a config that set it kept
    getting the built-in 5 attempts; it now gets the number it asked for.
  • An aborted lookup is no longer retried. Nothing classified
    OnlineLookupAbortedError, so it fell through to the retriable default
    and was replayed on the rate-limit schedule.

• Features

  • Comicbox warns when Metron's daily quota drops below 10%, stops starting
    new searches near the floor so the rest of the budget finishes comics that
    already matched, and stops cleanly at zero. Re-running is cheap — the
    response cache replays what was already fetched.
  • The end-of-run summary reports API cost: requests sent per endpoint,
    rejections, seconds spent pacing, and the remaining per-minute and daily
    budgets.
  • online.tuning.per_source.metron.rate_limit.per_minute works again, as a
    ceiling on the server-reported limit. Set it to share one token between
    several processes. Note that a config which already sets it was being
    warned it was ignored, and will now actually be held to it. per_day is
    still ignored and still says so.
  • --jobs is no longer capped at 20 for Metron runs. The cap bounded
    workers, which is not what Metron throttles; the gate bounds requests.
  • The outgoing User-Agent identifies the entry point, e.g.
    comicbox/5.1.0 (cli; jobs=8). OnlineSession(client_name=...) names an
    embedding application.
  • metron_requests_for_batch(comics, series) prices a batch the way one
    actually runs — each series resolved once, the rest of its comics at the
    warm price — instead of charging every comic a cold search. Pacing math
    now prefers the burst limit Metron reported over the documented default.

• Performance

  • A -j N batch resolves each series once instead of once per worker.
    Workers that want the same series wait for whoever is resolving it and
    then take the warm path, and the dispatcher hands the pool one file per
    series first, so its first N tasks are N different series rather than N
    copies of the same cold search.
  • Metron batches holding many comics from one series fetch that series'
    issue list once instead of once per comic, when Metron's reported issue
    count says that is actually cheaper. A 100-issue single-series run drops
    from about 200 requests to about 102.
  • Comic Vine tagging spends half the API budget it used to. simyan 4.1.0
    stops paginating past a short page, so each candidate volume's issue
    lookup costs one request against Comic Vine's 200-per-hour issues pool
    instead of two. That pool paces a Comic Vine run, so runs are faster as
    well as cheaper, and estimate_run() now projects the request count
    rather than a floor half its size.

• Dev

  • Require simyan >= 4.1.0.