Repository navigation
question(registry): should a built-in engine, reassembly or traceflow overridden by name be restorable through the public registrar? #1363
Description
This is a question and needs the owner's ruling; it may be by design.
Describe the behaviour
The public registrars accept an override of a built-in name, but refuse the built-in class when it is offered back, because they check the public Engine / Reassembly / TraceFlow (pcapkit/foundation/extraction.py:469) while the built-ins derive from the *Base classes. #513 and #1016 recorded this on purpose (_register_internal_* is the door for built-ins), so a restore needs the private method or a saved copy of the registry.
To Reproduce
register_extractor_engine('dpkt', MyEngine) works (with a RegistryWarning). register_extractor_engine('dpkt', DPKT) then raises RegistryError: engine must be an Engine subclass, not <class 'pcapkit.foundation.engines.dpkt.DPKT'>. The same holds for reassembly ipv4 / ipv6 / tcp and traceflow tcp.
Question
Is "an override of a built-in is permanent for the process" the intended contract? If so, register_engine's docstring could say so; if not, the public registrars could accept a class already registered under another name by the library, or register_* could offer an unregister / restore path.
Found by the round-trip audit of pcapkit/foundation. Run /tmp/rtissues/foundation-repro.py from the repo root (it prints pcapkit.__file__ first).
Found by the foundation round-trip audit (#1202). The repro was re-checked by an independent verifier, and I re-ran it on main at c9676c563.
Activity
JarryShaw commentedon Oct 7, 2026
Decision needed. #513 and #1016 made the public registrars refuse built-ins on purpose:
extraction.py:458-469records it in a comment and the docstring. The consequence is that onceregister_extractor_engine('dpkt', MyEngine)overrides a built-in, puttingDPKTback raisesRegistryError: engine must be an Engine subclass. The same holds for reassemblyipv4/ipv6/tcpand traceflowtcp.Options:
unregister_*/reset_*that restores the shipped default.I recommend (1). It is the smallest change, keeps #1016's guarantee for every foreign class, and makes the round trip override→restore exact.
JarryShaw commentedon Oct 7, 2026
okay take (1) then, and this will apply to all registrars.
JarryShaw commentedon Oct 7, 2026
Ruling recorded: option (1), applied to every registrar. A registrar accepts the shipped original for a name it shipped with, so an override can be undone; every other foreign class is still refused as #1016 requires. The rule covers the extraction engine, reassembly, traceflow and dumper registrars, and also the protocol and option registrars.
Sequencing:
foundation/extraction.py,foundation/traceflow/,foundation/registry/). No open PR edits them, so they go to a worker this tick.ipv4.py), fix(mh): keep zero-length PadN, reserved fields and BU/BA flag bits on rebuild #1345 (mh.py), fix(tcp)!: reserved bits, post-EOOL octets and the UTO granularity round-trip (#1211, #1218, #1219) #1348 (tcp.py) and fix(protocol): raw items in a make() list build a packet; str enum names resolve per argument #1352 (protocol.py). They join fix(registry): thirteen code-keyed registrars warn "already registered, overwriting" on re-registering the identical object #1362's sweep once those merge.fix(registry): foundation registrars restore the built-in a name ship…
JarryShaw commentedon Oct 8, 2026
The foundation half of this ruling has landed in #1376, which has merged. The protocol-registrar half rides on #1362. Blocked until #1362 is fixed, and #1362 itself waits on #1390.
fix(protocols): parser registrars warn only when the entry really cha…
JarryShaw commentedon Oct 8, 2026
The protocol-registrar half has landed in #1400, which merged: restoring a shipped entry now warns once and is accepted. What remains is the class-storing
registerinprotocols/protocol.py, which is #1364's code. Blocked until #1364 is fixed.JarryShaw commentedon Oct 8, 2026
Unblocked: #1364 is fixed by #1415 (merged at
199bfcef8). The class-storingregisterinprotocols/protocol.pynow keeps a re-registered stored descriptor, so the protocol-registrar half of this question is settled.What remains is the question itself, for the owner. It concerns the public
register_extractor_engine/_reassembly/_traceflowregistrars (foundation/extraction.py). Should offering a built-in class back after an override:RegistryWarning, the way fix(protocols): parser registrars warn only when the entry really changes (#1362) #1400 now treats protocol registrars?Moved to
needs: decision.JarryShaw commentedon Oct 8, 2026
Your recommendations?
JarryShaw commentedon Oct 8, 2026
Correction first: my comment above was stale. Option 2 is already on
main. #1376 (merged 2026-10-08) made the extractor registrars accept the built-in a name shipped with, #1400 did the same for the 14 parser registrars, and #1415 closed the last gap (#1364). Re-measured on949b0a900:register_extractor_engine('dpkt', MyEngine), then('dpkt', DPKT): accepted, and the registry holds the shippedModuleDescriptoragain.ipv4and traceflowtcp.register_extractor_engine('foo', DPKT)still raisesRegistryError. A built-in is only accepted back under the name it shipped with.tests/foundation/registry/test_restore_builtin_unit.py.Recommendation: close this as answered. Keep both behaviours:
Engine/Reassembly/TraceFlow, which carry the registration hook.One optional nit: restoring warns
already registered, overwriting, the same as an override. I'd keep it, because it does replace the override. If you'd rather it sayrestoring shipped entry, that is a one-line message change.JarryShaw commentedon Oct 8, 2026
all good. close it then.
JarryShaw commentedon Oct 8, 2026
Closing as answered, per the owner's ruling: the restore behaviour is already on
mainand both halves stay as they are. A built-in is accepted back under the name it shipped with (#1376, #1400, #1415); under any other name the public door still refuses it (#513, #1016). Thealready registered, overwritingwarning on a restore is kept as-is.