Every `Build()` method now requires a `context.Context`. This is so the rolodex knows where to resolve from when locating relative links. Without knowing where we are, there is no way to resolve anything. This new mechanism allows the model to recurse across as many files as required to locate references, without loosing track of where we are in the process.
Signed-off-by: quobix <dave@quobix.com>
This was brought up in chat, that the expectation is that all schemas are included, refs or not. This update fixes that for the index and adds in a new individual property to count just the references `GetAllReferenceSchemas `
Signed-off-by: Dave Shanley <dave@quobix.com>
In vacuum, a usecase was reported where an infinite loop occurred due to re-parsing the same reference over and over in a loop. It was re-creatable and it was because the loop happened before the index was ready.
This should be resolved now, at least for this use case. To be sure, I have included the specs as a new test.
https://github.com/daveshanley/vacuum/issues/268
Signed-off-by: Dave Shanley <dave@quobix.com>
Different race conditions popped up at different times, depending on how complex the nesting of indexes was. After running tests with `-race` I've knocked out all the concurrency issues being reported and now there are no race conditions reported by the tests. This should knock out all known race conditions with some targeted mutex locks.
Non breaking change.
The index runs async everywhere, it's kinda impossible to know which path with resolve first, so testing is hard. Sometimes a race condition is hit, well, it was. Now that map has a mutex on it.
Also fully fixed handling files with relative links. A basepath property has been added to the index configuration to allow a local root to be set when resolving files.
Added a full checkout test for digital ocean so that full remote and full local testing is performed.
Tests all passing, runs super fast, pulls in every single DigitalOcean spec and parses it. There may be some issues deeper down in the models, but for now high level tests all pass.
There is new code that looks for inline `schema` definitions and collects them, as well as a futher subset of references that are specifically `object` or `array` types. This is a breaking change because the `index.GetAllSchemas()` method now returns a slice of *Reference pointers, rather than a map of them. The original behavior of `GetAllSchemas()` has been replaced with `GetAllComponentSchemas()` which retains the signature.
Two new additional methods `GetAllInlineSchemas()` and `GetAllInlineSchemaObjects()`. The former will return everything in the spec marked with `schema` that is not a reference. The latter will return everything the former does, except filtered down by `object` or `array` types.
Also the index is now bubbled up through the document model. There isn't a simple way to get access to it, small non breaking change that attaches a reference to the index, returned by `DocumentModel`
This patch bumps up support for creating valid JSON paths from references. It addresses #48 and makes invalid specs, valid! This addresses issues #48 and #45 that report references cannot be found. Now `libopenapi` can support much more funky lookups.
If a remote reference is used that does not contain a component ID (http://something.com/somefile.json#/components/something). The index would throw an error. This is not the correct behavior as the raw file can be the root, thus no componentID is required.
The bulk of the code is expecting that componentID, so this fix will prepare any component-less remote reference, to be treated properly.
Signed-off-by: Dave Shanley <dave@quobix.com>
When paths are wrapped in an array, the index throws an error because it's expecting them to be supplied as a map. Now paths can be wrapped or unwrapped, the index won't care.
Swagger docs define securityDefinitions, however they are not tracked by the index, and vacuum needs to know about them, so now the index collects these references.
A single line that tries to read an HTTP response body and fails is very hard to test without mocking, and the mock does not add value to a single line of code to check for an error that can rarely ever be triggered. Going to settle for 99.9% for now.
Same functions will power every other model. Found some opprotunities to add control over circular references, when and how to fire errors and what not. All in all, some robust code additions will improve the library significantly.
Building out models that have circular references is tricky. Particularly when they are polymorphic ones. The resolver now searches the polymorphic references for circular references as well (which is returning much more valuable results now). The low level model is now sensitive to this and will avoid building out anything that is circular.
I think it should be pretty much bullet proof at the moment (until the next failure). Also added external index lookups from main index for external docs. found this bug after refining lookups in schemas.
Stripe API is slowest to build at about 1.2s, the depth of the schemas on that spec are insane. The only way to speed things up here is to NOT wait for threads to finish before returning the model, and this creates a broken guarantee as the model won't be finished building by the time its returned to the consumer. When running tests against this - they fail randomly. depending on which thread reaches the finish first, before trying to be read.
This isn't an option, so for now, this is as fast as she will get, which is pretty good. 5ms for petstore and about 250ms for k8s.
