fix: Populate the schema cache with local defs.json again

setup.py

@@ -40,6 +40,7 @@
             'scrapbook~=0.5.0',
             'jupyter',
             'graphviz',
+            'pytest-socket>=0.2',
         ]
     )
 )

src/pyhf/schema/__init__.py

@@ -72,6 +72,8 @@ def __call__(self, new_path: pathlib.Path):
             self (pyhf.schema.Schema): Returns itself (for contextlib management)
         """
         self.orig_path, variables.schemas = variables.schemas, new_path
+        self.orig_cache = dict(variables.SCHEMA_CACHE)
+        variables.SCHEMA_CACHE.clear()
         return self
 
     def __enter__(self):
@@ -85,6 +87,7 @@ def __exit__(self, *args, **kwargs):
             None
         """
         variables.schemas = self.orig_path
+        variables.SCHEMA_CACHE = self.orig_cache
 
     @property
     def path(self):

src/pyhf/schema/loader.py

@@ -39,3 +39,9 @@ def load_schema(schema_id: str):
             schema = json.load(json_schema)
             variables.SCHEMA_CACHE[schema['$id']] = schema
         return variables.SCHEMA_CACHE[schema['$id']]
+
+
+# pre-populate the cache to avoid network access
+# on first validation in standard usage
+# (not in pyhf.schema.variables to avoid circular imports)
+load_schema(f'{variables.SCHEMA_VERSION}/defs.json')

tests/test_schema.py

@@ -1,6 +1,9 @@
 import pyhf
 import pytest
 import json
+import importlib
+import sys
+from pytest_socket import socket_disabled  # noqa: F401
 
 
 @pytest.mark.parametrize('version', ['1.0.0'])
@@ -27,11 +30,20 @@ def test_schema_callable():
     assert callable(pyhf.schema)
 
 
-def test_schema_changeable(datadir, monkeypatch):
+@pytest.fixture
+def self_restoring_schema_globals():
+    old_path = pyhf.schema.path
+    old_cache = dict(pyhf.schema.variables.SCHEMA_CACHE)
+    yield old_path, old_cache
+    pyhf.schema(old_path)
+    pyhf.schema.variables.SCHEMA_CACHE = old_cache
+
+
+def test_schema_changeable(datadir, monkeypatch, self_restoring_schema_globals):
     monkeypatch.setattr(
         pyhf.schema.variables, 'schemas', pyhf.schema.variables.schemas, raising=True
     )
-    old_path = pyhf.schema.path
+    old_path, old_cache = self_restoring_schema_globals
     new_path = datadir / 'customschema'
 
     with pytest.raises(pyhf.exceptions.SchemaNotFound):
@@ -40,36 +52,47 @@ def test_schema_changeable(datadir, monkeypatch):
     pyhf.schema(new_path)
     assert old_path != pyhf.schema.path
     assert new_path == pyhf.schema.path
+    assert pyhf.schema.variables.SCHEMA_CACHE is not old_cache
+    assert len(pyhf.schema.variables.SCHEMA_CACHE) == 0
     assert pyhf.Workspace(json.load(open(new_path / 'custom.json')))
-    pyhf.schema(old_path)
+    assert len(pyhf.schema.variables.SCHEMA_CACHE) == 1
 
 
-def test_schema_changeable_context(datadir, monkeypatch):
+def test_schema_changeable_context(datadir, monkeypatch, self_restoring_schema_globals):
     monkeypatch.setattr(
         pyhf.schema.variables, 'schemas', pyhf.schema.variables.schemas, raising=True
     )
-    old_path = pyhf.schema.path
+    old_path, old_cache = self_restoring_schema_globals
     new_path = datadir / 'customschema'
 
     assert old_path == pyhf.schema.path
     with pyhf.schema(new_path):
         assert old_path != pyhf.schema.path
         assert new_path == pyhf.schema.path
+        assert pyhf.schema.variables.SCHEMA_CACHE is not old_cache
+        assert len(pyhf.schema.variables.SCHEMA_CACHE) == 0
         assert pyhf.Workspace(json.load(open(new_path / 'custom.json')))
+        assert len(pyhf.schema.variables.SCHEMA_CACHE) == 1
     assert old_path == pyhf.schema.path
+    assert old_cache == pyhf.schema.variables.SCHEMA_CACHE
 
 
-def test_schema_changeable_context_error(datadir, monkeypatch):
+def test_schema_changeable_context_error(
+    datadir, monkeypatch, self_restoring_schema_globals
+):
     monkeypatch.setattr(
         pyhf.schema.variables, 'schemas', pyhf.schema.variables.schemas, raising=True
     )
-    old_path = pyhf.schema.path
+    old_path, old_cache = self_restoring_schema_globals
     new_path = datadir / 'customschema'
 
     with pytest.raises(ZeroDivisionError):
         with pyhf.schema(new_path):
+            # this populates the current cache
+            pyhf.Workspace(json.load(open(new_path / 'custom.json')))
             raise ZeroDivisionError()
     assert old_path == pyhf.schema.path
+    assert old_cache == pyhf.schema.variables.SCHEMA_CACHE
 
 
 def test_no_channels():
@@ -567,3 +590,54 @@ def test_patchset_fail(datadir, patchset_file):
     patchset = json.load(open(datadir.joinpath(patchset_file)))
     with pytest.raises(pyhf.exceptions.InvalidSpecification):
         pyhf.schema.validate(patchset, 'patchset.json')
+
+
+@pytest.fixture
+def refresh_pyhf(monkeypatch):
+    global pyhf
+    modules_to_clear = [name for name in sys.modules if name.split('.')[0] == 'pyhf']
+    for module_name in modules_to_clear:
+        monkeypatch.delitem(sys.modules, module_name)
+    pyhf = importlib.import_module(pyhf.__name__)
+    return pyhf
+
+
+def test_defs_always_cached(
+    socket_disabled,  # noqa: F811
+    refresh_pyhf,  # ensure there is not a pre-existing cache hiding the issue
+):
+    """
+    Schema definitions should always be loaded from the local files and cached at first import.
+
+    Otherwise using pyhf in contexts where the jsonschema.RefResolver cannot lookup the definition by the schema-id,
+    it will crash (e.g. a cluster node without network access).
+    """
+    pyhf = refresh_pyhf
+    spec = {
+        'channels': [
+            {
+                'name': 'singlechannel',
+                'samples': [
+                    {
+                        'name': 'signal',
+                        'data': [10],
+                        'modifiers': [
+                            {'name': 'mu', 'type': 'normfactor', 'data': None}
+                        ],
+                    },
+                    {
+                        'name': 'background',
+                        'data': [20],
+                        'modifiers': [
+                            {
+                                'name': 'uncorr_bkguncrt',
+                                'type': 'shapesys',
+                                'data': [30],
+                            }
+                        ],
+                    },
+                ],
+            }
+        ]
+    }
+    pyhf.schema.validate(spec, 'model.json')  # may try to access network and fail