Support the HTCondor v2 python bindings

[ikrommyd]

🤖 AI text below 🤖

HTCondor 25 removed the v1 python bindings (HTCONDOR-3240), so import htcondor fails on current coffea images — e.g. coffeateam/coffea-0.7-almalinux9:0.7.31-py3.11, which ships htcondor 25.11.0. The distribution is still named htcondor, so pip show htcondor still finds it; only the module was renamed to htcondor2.

This ports the package rather than pinning, so it works on both. Alternative to #45.

Three v2 differences mattered:

• import htcondor → prefer htcondor2, fall back to htcondor.
• HTCondorIOError is gone, along with every other HTCondorException subclass. Aliased to the surviving base class.
• Submit.jobs() raises NotImplementedError in v2, and Schedd.spool() now takes the SubmitResult instead of a list of job ads. Both are used together in LPCCondorJob.start, so they are wrapped in a spool() helper in schedd.py.

That third one is the reason this is more than an import fix — without it the package imports cleanly and then fails on every job submission.

Everything else works unchanged: param, Collector, Schedd.query/act, Submit, JobAction, and the AdTypes/DaemonTypes enums, which htcondor2 keeps as explicit back-compat aliases. xquery() isn't used anywhere, so there is no query-loop to rewrite. No setup.cfg change is needed since the distribution name is unchanged.

Verified

Against real conda-forge python-htcondor builds on py3.11:

version	modules present	selected	checked
23.0.27	htcondor only	v1	real HTCondorIOError subclass preserved; Submit.jobs() works; spool() gets a list — i.e. pre-PR behaviour
23.10.27	both	v2	module selection
24.0.21	both	v2	full suite passes
25.11.0	htcondor2 only	v2	from lpcjobqueue import LPCCondorCluster OK; every touched attribute resolves; Submit.jobs() raises NotImplementedError; Schedd.spool is (result: SubmitResult); except HTCondorIOError catches HTCondorException

Two things worth knowing:

• htcondor2 exists from 23.10 onward, so this moves 23.10/24 users onto the v2 bindings too, not just 25 users. That is deliberate — v1 is the Boost.Python build that won't get wheels for newer Pythons — but it is a behaviour change for people who are working fine today. Happy to invert the preference to v1-first if you'd rather only move users who have no choice.
• Not covered: a live submit/spool round trip against a real schedd, and acquire_schedd() against a live collector. Those need an LPC node, so a smoke test before merging would be worthwhile — the spool change above is exactly the path I could not exercise end to end.

One note for maintainers

The HTCondorIOError alias widens the three except clauses in cluster.py to catch any HTCondorException. All three are schedd-communication retry paths so this is mostly benign, but non-IO HTCondor errors will now be retried rather than propagating. If you'd prefer it explicit, I can swap the alias for a SCHEDD_RETRY_ERRORS tuple in schedd.py and use that in the except clauses instead.