Details
-
Type:
Story
-
Status: Done
-
Resolution: Done
-
Fix Version/s: None
-
Component/s: afw
-
Labels:
-
Story Points:1
-
Team:Architecture
Description
If astropy.units fails to import and a raises an exception (eg if warnings are converted to errors) when being imported via Schema.cc during import of lsst.afw.image an abort signal is received:
terminating with uncaught exception of type pybind11::error_already_set: SystemError: <built-in method __contains__ of dict object at 0x10049b900> returned a result with an error set
|
Stepping through schema.cc the line triggering this is:
* thread #1, queue = 'com.apple.main-thread', stop reason = breakpoint 2.1
|
frame #0: 0x0000000116e3d53d schema.so`void lsst::afw::table::(anonymous namespace)::declareSchemaType<unsigned char>(mod=0x00007ffeefbf33c0) at schema.cc:187 [opt]
|
184 std::string suffix = FieldBase<T>::getTypeString();
|
185 py::str pySuffix(suffix);
|
186
|
-> 187 py::object astropyUnit = py::module::import("astropy.units").attr("Unit");
|
188
|
189 // FieldBase
|
190 PyFieldBase<T> clsFieldBase(mod, ("FieldBase" + suffix).c_str());
|
|
|
* frame #0: 0x0000000116e3d53d schema.so`void lsst::afw::table::(anonymous namespace)::declareSchemaType<unsigned char>(mod=0x00007ffeefbf33c0) at schema.cc:187 [opt]
|
frame #1: 0x0000000116e3d143 schema.so`lsst::afw::table::(anonymous namespace)::pybind11_init_schema(mod=0x00007ffeefbf33c0) at schema.cc:398 [opt]
|
frame #2: 0x0000000116e3cf5d schema.so`::PyInit_schema() at schema.cc:390 [opt]
|
|
Something in the pybind11 layer is failing to catch the python exception and forward it on.
Depending on your system this can be reproduced with:
import warnings |
warnings.simplefilter("error") |
import lsst.afw.image |
Given the reference to schema.cc and _contains_ I initially thought this was related to DM-15406 and there is some commentary there.
Attachments
Issue Links
Activity
Good news is that pybind11 master does not SIGABRT – you get the annoyingly incomplete rewritten exception text. I'll file a ticket anyhow to see if I can stop the exception rewriting. In the mean time I can merge this for the older pybind11 we have.
Christopher Waters and I just ran into another error_already_set terminate in a new branch for DM-14980, so it'd definitely be nice to get this patch onto (LSST) master to make it easy to test if it fixes that problem as well.
Merged. Jenkins passed.
(All of this assumes that none of us actually understand why adding e.restore() prevents a C++ exception from propagating up to terminate - if you or Paul Price does and I'm just not seeing it, please speak up.)