A user who calls `StreamInvoke` probably expects the `AsyncIterable`
that is returned to gracefully terminate. This is currently not the
case.
Where does something like this go wrong? A better question might be
where any of this went right, because several days later, after
wandering into civilization from the great Wilderness of Bugs, I must
confess that I've forgotten if any of it had.
`AsyncIterable` is a pull-based API. `for await (...)` will continuously
call `next` ("pull") on the underlying `AsyncIterator` until the
iterable is exhausted. But, gRPC's streaming-return API is _push_ based.
That is to say, when a streaming RPC is called, data is provided by
callback on the stream object, like:
call.on("data", (thing: any) => {... do thing ...});
Our goal in `StreamInvoke` is to convert the push-based gRPC routines
into the pull-based `AsyncIterable` retrun type. You may remember your
CS theory this is one of those annoying "fundamental mismatches" in
abstraction. So we're off to a good start.
Until this point, we've depended on a library,
`callback-to-async-iterator` to handle the details of being this bridge.
Our trusting nature and innocent charm has mislead us. This library is
not worthy of our trust. Instead of doing what we'd like it to do, it
returns (in our case) an `AsyncIterable` that will never complete.
Yes,, this `AsyncIterable` will patiently wait for eternity, which
honestly is kind of poetic when you sit down in a nice bath and think
about that fun time you considered eating your computer instead of
finishing this idiotic bug.
Indeed, this is the sort of bug that you wonder where it even comes
from. Our query libraries? Why aren't these `finally` blocks executing?
Is our language host terminating early? Is gRPC angry at me, and just
passive-aggrssively not servicing some of my requests? Oh god I've been
up for 48 hours, why is that wallpaper starting to move? And by the way,
a fun interlude to take in an otherwise very productive week is to try
to understand the gRPC streaming node client, which is code-gen'd, but
which also takes the liberty of generating itself at runtime, so that
gRPC is code-gen'ing a code-gen routine, which makes the whole thing
un-introspectable, un-debuggable, and un-knowable. That's fine, I didn't
need to understand any of this anyway, thanks friends.
But we've come out the other side knowing that the weak link in this
very sorry chain of incredibly weak links, is this dependency.
This commit removes this dependency for a better monster: the one we
know.
It is at this time that I'd like to announce that I am quitting my job
at Pulumi. I thank you all for the good times, but mostly, for taking
this code over for me.
These changes restore a more-correct version of the behavior that was
disabled with #3014. The original implementation of this behavior was
done in the SDKs, which do not have access to the complete inputs for a
resource (in particular, default values filled in by the provider during
`Check` are not exposed to the SDK). This lack of information meant that
the resolved output values could disagree with the typings present in
a provider SDK. Exacerbating this problem was the fact that unknown
values were dropped entirely, causing `undefined` values to appear in
unexpected places.
By doing this in the engine and allowing unknown values to be
represented in a first-class manner in the SDK, we can attack both of
these issues.
Although this behavior is not _strictly_ consistent with respect to the
resource model--in an update, a resource's output properties will come
from its provider and may differ from its input properties--this
behavior was present in the product for a fairly long time without
significant issues. In the future, we may be able to improve the
accuracy of resource outputs during a preview by allowing the provider
to dry-run CRUD operations and return partially-known values where
possible.
These changes also introduce new APIs in the Node and Python SDKs
that work with unknown values in a first-class fashion:
- A new parameter to the `apply` function that indicates that the
callback should be run even if the result of the apply contains
unknown values
- `containsUnknowns` and `isUnknown`, which return true if a value
either contains nested unknown values or is exactly an unknown value
- The `Unknown` type, which represents unknown values
The primary use case for these APIs is to allow nested, properties with
known values to be accessed via the lifted property accessor even when
the containing property is not fully know. A common example of this
pattern is the `metadata.name` property of a Kubernetes `Namespace`
object: while other properties of the `metadata` bag may be unknown,
`name` is often known. These APIs allow `ns.metadata.name` to return a
known value in this case.
In order to avoid exposing downlevel SDKs to unknown values--a change
which could break user code by exposing it to unexpected values--a
language SDK must indicate whether or not it supports first-class
unknown values as part of each `RegisterResourceRequest`.
These changes also allow us to avoid breaking user code with the new
behavior introduced by the prior commit.
Fixes#3190.
After creating an initial Pulumi .NET project, it will be natural for some folks (who are unfamiliar with Pulumi) to try to run it via `dotnet run`, as that's how you'd typically run a .NET Core program. Doing so today fails with:
```
Unhandled exception. System.InvalidOperationException: Environment did not contain: PULUMI_MONITOR
at Pulumi.Deployment..ctor()
at Pulumi.Deployment.RunAsync(Func`1 func)
at Pulumi.Deployment.RunAsync(Func`1 func)
at Pulumi.Deployment.RunAsync(Action action)
at Program.Main() in /Users/user/temp/quickstart/Program.cs:line 9
at Program.<Main>()
```
Instead, provide a more descriptive error message indicating that the pulumi CLI should be used to run the program. We return the same error as we do for Node.js and Python.
```
Unhandled exception. System.InvalidOperationException: Program run without the Pulumi engine available; re-run using the `pulumi` CLI
at Pulumi.Deployment..ctor()
at Pulumi.Deployment.RunAsync(Func`1 func)
at Pulumi.Deployment.RunAsync(Func`1 func)
at Pulumi.Deployment.RunAsync(Action action)
at Program.Main() in /Users/user/temp/quickstart/Program.cs:line 9
at Program.<Main>()
```
The @pulumi/pulumi TypScript SDK exposes `streamInvoke`, which returns a
(potentially infinite) stream of responses. This currently is _assumed_
to be infinite, in that there is no way to signal cancellation, and
prevents Pulumi from being able to clean up when we're finished using
the results of the `streamInvoke`.
This commit will introduce a `StreamInvokeResult` type, which is an
`AsyncIterable` that also exposes a `cancel` function, whih does just
this.
Use it like this:
// `streamInvoke` to retrieve all updates to any `Deployment`, enumerate 0
// updates from the stream, then `cancel` giving the Kubernetes provider to
// clean up and close gracefully.
const deployments = await streamInvoke("kubernetes:kubernetes:watch", {
group: "apps", version: "v1", kind: "Deployment",
break;
});
deployments.cancel();
If a stack output includes a `Resource`, we will as of a recent change
always show the output diff, but this diff will potentially include
unknowns, leading to spurious output like:
```
+ namePrefix : output<string>
```
These changes supress these diffs by adding a special key to the POJO
we generate for resources *during preview only* that indicates that the
POJO represents a Pulumi resource, then stripping all adds of unknown
values from diffs for objects marked with that key.
Fixes#3314.
Adds Python support for resource transformations aligned with the existing NodeJS support in #3174.
This PR also moves processing of transformations to earlier in the resource construction process (for both NodeJS and Python) to ensure that invariants established in the constructor cannot be violated by transformations. This change can technically be a breaking change, but given that (a) the transformations features was just released in 1.3.0 and (b) the cases where this is a breaking change are uncommon and unlikely to have been reliable anyway - it feels like a change we should make now.
Fixes#3283.
These changes restore a more-correct version of the behavior that was
disabled with #3014. The original implementation of this behavior was
done in the SDKs, which do not have access to the complete inputs for a
resource (in particular, default values filled in by the provider during
`Check` are not exposed to the SDK). This lack of information meant that
the resolved output values could disagree with the typings present in
a provider SDK. Exacerbating this problem was the fact that unknown
values were dropped entirely, causing `undefined` values to appear in
unexpected places.
By doing this in the engine and allowing unknown values to be
represented in a first-class manner in the SDK, we can attack both of
these issues.
Although this behavior is not _strictly_ consistent with respect to the
resource model--in an update, a resource's output properties will come
from its provider and may differ from its input properties--this
behavior was present in the product for a fairly long time without
significant issues. In the future, we may be able to improve the
accuracy of resource outputs during a preview by allowing the provider
to dry-run CRUD operations and return partially-known values where
possible.
These changes also introduce new APIs in the Node and Python SDKs
that work with unknown values in a first-class fashion:
- A new parameter to the `apply` function that indicates that the
callback should be run even if the result of the apply contains
unknown values
- `containsUnknowns` and `isUnknown`, which return true if a value
either contains nested unknown values or is exactly an unknown value
- The `Unknown` type, which represents unknown values
The primary use case for these APIs is to allow nested, properties with
known values to be accessed via the lifted property accessor even when
the containing property is not fully know. A common example of this
pattern is the `metadata.name` property of a Kubernetes `Namespace`
object: while other properties of the `metadata` bag may be unknown,
`name` is often known. These APIs allow `ns.metadata.name` to return a
known value in this case.
In order to avoid exposing downlevel SDKs to unknown values--a change
which could break user code by exposing it to unexpected values--a
language SDK must indicate whether or not it supports first-class
unknown values as part of each `RegisterResourceRequest`.
These changes also allow us to avoid breaking user code with the new
behavior introduced by the prior commit.
Fixes#3190.
Adds the ability to provide `transformations` to modify the properties and resource options that will be used for any child resource of a component or stack.
This offers an "escape hatch" to modify the behaviour of a component by peeking behind it's abstraction. For example, it can be used to add a resource option (`additionalSecretOutputs`, `aliases`, `protect`, etc.) to a specific known child of a component, or to modify some input property to a child resource if the component does not (yet) expose the ability to control that input directly. It could also be used for more interesting scenarios - such as:
1. Automatically applying tags to all resources that support them in a stack (or component)
2. Injecting real dependencies between stringly-referenced resources in a Helm Chart
3. Injecting explicit names using a preferred naming convention across all resources in a stack
4. Injecting `import` onto all resources by doing a lookup into a name=>id mapping
Because this feature makes it possible to peek behind a component abstraction, it must be used with care in cases where the component is versioned independently of the use of transformations. Also, this can result in "spooky action at a distance", so should be used judiciously. That said - this can be used as an escape hatch to unblock a wide variety of common use cases without waiting on changes to be made in a component implementation.
Each transformation is passed the `resource`, `name`, `type`, `props` and `opts` that are passed into the `Resource` constructor for any resource descended from the resource that has the transformation applied. The transformation callback can optionally return alternate versions of the `props` and `opts` to be used in place of the original values provided to the resource constructor.
Fixes#2068.
* Fix some tracing issues.
- Add endpoints for `startUpdate` and `postEngineEventsBatch` so that
spans for these invocations have proper names
- Inject a tracing span when walking a plan so that resource operations
are properly parented
- When handling gRPC calls, inject a tracing span into the call's
metadata if no span is already present so that resource monitor and
engine spans are properly parented
- Do not trace client gRPC invocations of the empty method so that these
calls (which are used to determine server availability) do not muddy
the trace. Note that I tried parenting these spans appropriately, but
doing so broke the trace entirely.
With these changes, the only unparented span in a typical Pulumi
invocation is a single call to `getUser`. This span is unparented
because that call does not have a context available. Plumbing a context
into that particular call is surprisingly tricky, as it is often called
by other context-less functions.
* Make tracing support more flexible.
- Add support for writing trace data to a local file using Appdash
- Add support for viewing Appdash traces via the CLI
