With the clean up of deployment options (#16357), it is no longer
necessary for `Step.Apply` methods to accept a `preview bool` parameter.
This information is already present on the receiver's `deployment`
property and removing the argument a. allows us to consolidate the
surface area of the API, hopefully preventing bugs and b. further
improve consistency around the use of "preview" (referring to the
`preview` action) and "dry run" (referring to any part of an action that
does not perform modifications, such as showing what will happen during
a `refresh` or `destroy`).
Presently, the behaviour of diffing during refresh steps is incomplete,
returning only an "output diff" that presents the changes in outputs.
This commit changes refresh steps so that:
* they compute a diff similar to the one that would be computed if a
`preview` were run immediately after the refresh, which is more
typically what users expect and want; and
* `IgnoreChanges` resource options are respected when performing the new
desired-state diffs, so that property additions or changes reported by a
refresh can be ignored.
In particular, `IgnoreChanges` can now be used to acknowledge that part
or all of a resource may change in the provider, but the user is OK with
this and doesn't want to be notified about it during a refresh.
Importantly, this means that the diff won't be reported, but also that
the changes won't be applied to state.
The implementation covers the following:
* A diff is computed using the inputs from the program and then
inverting the result, since in the case of a refresh the diff is being
driven by the provider side and not the program. This doesn't change
what is stored back into the state, but it does produce a diff that is
more aligned with the "true changes to the desired state".
* `IgnoreChanges` resource options are now stored in state, so that this
information can be used in refresh operations that do not have access
to/run the program.
* In the context of a refresh operation, `IgnoreChanges` applies to
*both* input and output properties. This differs from the behaviour of a
normal update operation, where `IgnoreChanges` only considers input
properties.
* The special `"*"` value for `IgnoreChanges` can be used to ignore all
properties. It _also_ ignores the case where the resource cannot be
found in the provider, and instead keeps the resource intact in state
with its existing input and output properties.
Because the program is not run for refresh operations, `IgnoreChanges`
options must be applied separately before a refresh takes place. This
can be accomplished using e.g. a `pulumi up` that applies the options
prior to a refresh. We should investigate perhaps providing a `pulumi
state set ...`-like CLI to make these sorts of changes directly to a
state.
For use cases relying on the legacy refresh diff provider, the
`PULUMI_USE_LEGACY_REFRESH_DIFF` environment variable can be set, which
will disable desired-state diff computation. We only need to perform
checks in `RefreshStep.{ResultOp,Apply}`, since downstream code will
work correctly based on the presence or absence of a `DetailedDiff` in
the step.
### Notes
- https://github.com/pulumi/pulumi/issues/16144 affects some of these
cases - though its technically orthogonal
- https://github.com/pulumi/pulumi/issues/11279 is another technically
orthogonal issue that many providers (at least TFBridge ones) - do not
report back changes to input properties on Read when the input property
(or property path) was missing on the inputs. This is again technically
orthogonal - but leads to cases that appear "wrong" in terms of what is
stored back into the state still - though the same as before this
change.
- Azure Native doesn't seem to handle `ignoreChanges` passed to Diff, so
the ability to ignore changes on refresh doesn't currently work for
Azure Native.
### Fixes
* Fixes#16072
* Fixes#16278
* Fixes#16334
* Not quite #12346, but likely replaces the need for that
Co-authored-by: Will Jones <will@sacharissa.co.uk>
# Description
There are a number of parts of the deployment process that require
context about and configuration for the operation being executed. For
instance:
* Source evaluation -- evaluating programs in order to emit resource
registrations
* Step generation -- processing resource registrations in order to
generate steps (create this, update that, delete the other, etc.)
* Step execution -- executing steps in order to action a deployment.
Presently, these pieces all take some form of `Options` struct or pass
explicit arguments. This is problematic for a couple of reasons:
* It could be possible for different parts of the codebase to end up
operating in different contexts/with different configurations, whether
due to different values being passed explicitly or due to missed
copying/instantiation.
* Some parts need less context/configuration than others, but still
accept full `Options`, making it hard to discern what information is
actually necessary in any given part of the process.
This commit attempts to clean things up by moving deployment options
directly into the `Deployment` itself. Since step generation and
execution already refer to a `Deployment`, they get a consistent view of
the options for free. For source evaluation, we introduce an
`EvalSourceOptions` struct for configuring just the options necessary
there. At the top level, the engine configures a single set of options
to flow through the deployment steps later on.
As part of this work, a few other things have been changed:
* Preview/dry-run parameters have been incorporated into options. This
lets up lop off another argument and mitigate a bit of "boolean
blindness". We don't appear to flip this flag within a deployment
process (indeed, all options seem to be immutable) and so having it as a
separate flag doesn't seem to buy us anything.
* Several methods representing parts of the deployment process have lost
arguments in favour of state that is already being carried on (or can be
carried on) their receiver. For instance, `deployment.run` no longer
takes actions or preview configuration. While doing so means that a
`deployment` could be run multiple times with different actions/preview
arguments, we don't currently exploit this fact anywhere, so moving this
state to the point of construction both simplifies things considerably
and reduces the possibility for error (e.g. passing different values of
`preview` when instantiating a `deployment` and subsequently calling
`run`).
* Event handlers have been split out of the options object and attached
to `Deployment` separately. This means we can talk about options at a
higher level without having to `nil` out/worry about this field and
mutate it correctly later on.
* Options are no longer mutated during deployment. Presently there
appears to be only one case of this -- when handling `ContinueOnError`
in the presence of `IgnoreChanges` (e.g. when performing a refresh).
This case has been refactored so that the mutation is no longer
necessary.
# Notes
* This change is in preparation for #16146, where we'd like to add an
environment variable to control behaviour and having a single unified
`Options` struct would make it easier to pass this configuration down
with introducing (more) global state into deployments. Indeed, this
change should make it easier to factor global state into `Options` so
that it can be controlled and tested more easily/is less susceptible to
bugs, race conditions, etc.
* I've tweaked/extended some comments while I'm here and have learned
things the hard way (e.g. `Refresh` vs `isRefresh`). Feedback welcome on
this if we'd rather not conflate.
* This change does mean that if in future we wanted e.g. to be able to
run a `Deployment` in multiple different ways with multiple sets of
actions, we'd have to refactor. Pushing state to the point of object
construction reduces the flexibility of the code. However, since we are
not presently using that flexibility (nor is there an obvious [to my
mind] use case in the near future), this seems like a good trade-off to
guard against bugs/make it simpler to move that state around.
* I've left some other review comments in the code around
questions/changes that might be a bad idea; happy to receive feedback on
it all though!
covers
- step.go
- step_executor.go
- step_generator.go
- import.go
They all depend on an added field to steps that use providers in
`Apply()`
# Includes changes to a non-test file: step.go
steps query the deployment for providers. There is not a straightforward
way of mocking a provider for a step. I've added a field called
`provider plugin.Provider` to steps that use providers to be used
instead of querying the Deployment with getProvider().
This approach aims to minimize the cognitive complexity and potential
for errors in comparison to the branching alternative due to the
behavior of `:=` and assigning the value to an existing value while also
defining a new variable err.
```diff
- prov, err := getProvider(s, s.provider)
- if err != nil {
- return resource.StatusOK, nil, err
+ prov := s.provider
+ if prov == nil {
+ var err error
+ prov, err = getProvider(s)
+ if err != nil {
+ return resource.StatusOK, nil, err
+ }
}
```