Field validator regression on dependent fields

[rlupton20]

#497 found a regression in validator function of fields in 2.0.1 dependent on prior fields. #499 purported to both fix this, and add tests to prevent future regressions.

I've produced a similar issue for dependent fields in 2.1.0. The docs here
https://schematics.readthedocs.io/en/latest/usage/validation.html#model-level-validation
suggest this should be handled.

from schematics.models import Model
from schematics.types import StringType
from schematics.exceptions import ValidationError
import schematics

class TestModel(Model):
    field1 = StringType(required=True)
    field2 = StringType()

    def validate_field2(self, data, value):
        if data['field1'] == "" and data['field2'] == "":
            raise ValidationError("Invalid values")
        return field2


print("Schematics version:", schematics.__version__)

TestModel({
    'field2': 'foo'
}).validate()

produces

Schematics version: 2.1.0
Traceback (most recent call last):
  File "schematics_test.py", line 18, in <module>
    'field2': 'foo'
  File "/usr/local/lib/python3.7/site-packages/schematics/models.py", line 258, in validate
    partial=partial, convert=convert, app_data=app_data, **kwargs)
  File "/usr/local/lib/python3.7/site-packages/schematics/models.py", line 299, in _convert
    return func(self._schema, self, raw_data=raw_data, oo=True, context=context, **kwargs)
  File "/usr/local/lib/python3.7/site-packages/schematics/validate.py", line 64, in validate
    errors.update(_validate_model(schema, mutable, data, context))
  File "/usr/local/lib/python3.7/site-packages/schematics/validate.py", line 95, in _validate_model
    schema._validator_functions[field_name](mutable, data, value, context)
  File "/usr/local/lib/python3.7/site-packages/schematics/validate.py", line 128, in newfunc
    return func(*args, **kwargs)
  File "schematics_test.py", line 11, in validate_field2
    if data['field1'] == "":
KeyError: 'field1'

Tested in the python:3.7 docker container - just pip install schematics and run the above to reproduce.

[rlupton20]

Actually what I've found is slightly different (i.e. not a re-regression), but the docs suggest should be fine. I'll update. It looks like it's all in the same vein.

[rlupton20]

The problem seems to be in this logic in validate.py

    errors = {}               
    try:                             
        data = import_loop(schema, mutable, raw_data, trusted_data=trusted_data,
            context=context, **kwargs)
    except DataError as exc:            
        errors = dict(exc.errors)
        data = exc.partial_data
                       
    errors.update(_validate_model(schema, mutable, data, context))

    if errors:
        raise DataError(errors, data)

In the above case we do populate error dictionary with an error to say field1 isn't set, but run _validate_model for field2 anyway, despite the fact there may be some dependency on something we already errored on.

Some ideas

• error before running the _validate_model functions if some data doesn't have the required properties. Then we lose having all errors at once.
• Become aware of dependencies in the _validate_models function. Perhaps handle failed data key lookups by cross-referencing with the current error dictionary? It all seems a bit gross.