qvd2parquet v2.1.0

The quality gate stops being the slow, silent, uninterruptible part of a wide
conversion. Everything here follows from running 2.0.0 on the SAP extract that
prompted it: the gate is now the default, and on a 20.6M-row, 213-column file
it ran for about twenty minutes at 11% CPU, printing nothing and ignoring
Ctrl-C.

Added

• The quality gate reads the output back in parallel, splitting it across
  workers by row group the same way decoding is split by chunk. It was entirely
  single-threaded, which on a wide file left the machine idle for minutes at
  the end of every run: on a 213-column, 1M-row fixture the full gate takes
  60.9s single-threaded and 9.3s at eight workers, cutting the whole run from
  4.7x the conversion to 2.6x. Row groups are independent, and both the metrics
  and the fingerprint merge in any order -- which is what already let parallel
  decoding validate without reordering -- so the verdict does not depend on the
  worker count. Each worker opens its own handle, for the same reason the
  decode workers do.

• The quality gate reports progress on the --progress cadence, which is on by
  default. It previously printed nothing until it finished, so a run that was
  working looked like one that had hung.

• Ctrl-C and SIGTERM now shut down gracefully and report themselves as what
  they are. A cancelled run stops at the next chunk boundary, drains what is in
  flight, removes the temporary output and exits with a new code 7.

  It previously exited 4, input error, with wrote 234725 rows but the header declares 1000000 -- a stopped run has written fewer rows than the
  header declares, which is exactly what a truncated input looks like, so
  pressing Ctrl-C told the user their QVD was corrupt. A cancelled quality gate
  likewise reported a gate failure, as though the output had not matched its
  input, when nobody had finished looking.

  A temporary output that cannot be deleted is now reported and named, instead
  of the failure being discarded and a partial Parquet file left sitting beside
  the real one. The delete is retried briefly first: Windows refuses to remove
  a file while any handle is open, and a virus scanner or the search indexer
  routinely holds one for a moment on a file just written, so the first attempt
  can fail on a file that is about to be perfectly deletable.

  In batch mode the files not yet started are recorded as cancelled too, rather
  than carrying a bare context.Canceled that mapped to the input-error code
  and reported unattempted files as unreadable ones. A cancelled batch
  outranks any individual file's verdict in the summary, since the rest were
  never tried.

  The signal handler is written against an explicit channel rather than
  signal.NotifyContext, whose stop function cancels the context as well as
  unregistering the handler: a goroutine waiting on Done cannot tell a real
  signal from the deferred cleanup of a successful run, and announced a
  cancellation on 37 of 40 successful conversions.

  The quality gate also honours cancellation at all now: it ran on
  context.Background(), so Ctrl-C during the read-back did nothing whatsoever
  -- on a wide file, minutes of a signal being ignored. And because
  signal.NotifyContext keeps swallowing signals once it has fired, a second
  Ctrl-C did nothing either. The first signal now restores the default handler
  and says so, so an impatient second one stops the process outright.

Install

Download the archive for your platform below, unpack it, and put qvd2parquet
on your PATH. Verify the download against SHA256SUMS:

shasum -a 256 -c SHA256SUMS --ignore-missing

Binaries are pure Go and statically linked, so they need no runtime
dependencies: Linux (amd64, arm64), Windows (amd64, arm64) and macOS (amd64,
arm64).