This change improves our output formatting by generally adding
fewer prefixes. As shown in pulumi/pulumi#359, we were being
excessively verbose in many places, including prefixing every
console.out with "langhost[nodejs].stdout: ", displaying full
stack traces for simple errors like missing configuration, etc.
Overall, this change includes the following:
* Don't prefix stdout and stderr output from the program, other
than the standard "info:" prefix. I experimented with various
schemes here, but they all felt gratuitous. Simply emitting
the output seems fine, especially as it's closer to what would
happen if you just ran the program under node.
* Do NOT make writes to stderr fail the plan/deploy. Previously
we assumed that any console.errors, for instance, meant that
the overall program should fail. This simply isn't how stderr
is treated generally and meant you couldn't use certain
logging techniques and libraries, among other things.
* Do make sure that stderr writes in the program end up going to
stderr in the Pulumi CLI output, however, so that redirection
works as it should. This required a new Infoerr log level.
* Make a small fix to the planning logic so we don't attempt to
print the summary if an error occurs.
* Finally, add a new error type, RunError, that when thrown and
uncaught does not result in a full stack trace being printed.
Anyone can use this, however, we currently use it for config
errors so that we can terminate with a pretty error message,
rather than the monstrosity shown in pulumi/pulumi#359.
As part of pulumi/pulumi-fabric#331, we've been exploring just using
undefined to indicate that a property value is absent during planning.
We also considered blocking the message loop to simplify the overall
programming model, so that all asynchrony is hidden.
It turns out ThereBeDragons 🐲 anytime you try to block the
message loop. So, we aren't quite sure about that bit.
But the part we are convicted about is that this Computed/Property
model is far too complex. Furthermore, it's very close to promises, and
yet frustratingly so far away. Indeed, the original thinking in
pulumi/pulumi-fabric#271 was simply to use promises, but we wanted to
encourage dataflow styles, rather than control flow. But we muddied up
our thinking by worrying about awaiting a promise that would never resolve.
It turns out we can achieve a middle ground: resolve planning promises to
undefined, so that they don't lead to hangs, but still use promises so
that asynchrony is explicit in the system. This also avoids blocking the
message loop. Who knows, this may actually be a fine final destination.
There's a fair bit of clean up in here, but the meat is:
* Allocate the language runtime gRPC client connection on the
goroutine that will use it; this eliminates race conditions.
* The biggie: there *appears* to be a bug in gRPC's implementation
on Linux, where it doesn't implement WaitForReady properly. The
behavior I'm observing is that RPC calls will not retry as they
are supposed to, but will instead spuriously fail during the RPC
startup. To work around this, I've added manual retry logic in
the shared plugin creation function so that we won't even try
to use the client connection until it is in a well-known state.
pulumi/pulumi-fabric#337 tracks getting to the bottom of this and,
ideally, removing the work around.
The other minor things are:
* Separate run.js into its own module, so it doesn't include
index.js and do a bunch of random stuff it shouldn't be doing.
* Allow run.js to be invoked without a --monitor. This makes
testing just the run part of invocation easier (including
config, which turned out to be super useful as I was debugging).
* Tidy up some messages.
We have an issue in the runtime right now where we serialize closures
asynchronously, meaning we make it possible to form cycles between
resource graphs (something that ought to be impossible in our model,
where resources are "immutable" after creation and cannot form cycles).
Let me tell you a tale of debugging this ...
Well, no, let's not do that. But thankfully I've left behind some
little utilities that might make debugging such a thing easier down
the road. Namely:
* By default, most of our core runtime promises leverage a leak handler
that will log an error message should the process exit with certain
critical unresolved promises. This error message will include some
handy context (like whether it was an input promise) as well as a
stack trace for its point of creation.
* Optionally, with a flag in runtime/debuggable.ts, you may wire up
a hang detector, for situations where we may want to detect this
situation sooner than process exit, using the regular message loop.
This uses a defined timeout, prints the same diagnostics as the
leak detector when a hang is detected, and is disabled by default.
This change makes a few simplifications to how properties are exposed in
the system, mostly in the name of usability, but also to feel a bit more
like "idiomatic JavaScript". Namely:
* Rename `then` to `mapValue`. This hopefully helps to suggest that this
is meant for a dataflow style of programming.
* Move Property<T> into the runtime module, and remove PropertyState<T>,
collapsing back down to a single type. This also eliminates some of the
messy internal runtime casting, accessing of internal members, etc.
* Export a Computed<T> interface from the root of the module. This is
the entirety of the public-facing surface area for properties, and
exposes that single `mapValue` member function. The internal runtime
logic understands how to handle Property<T>s specifically in addition
to Computed<T>s more generally (in case someone writes their own).
This changes a few aspects of resource properties:
* Move all runtime-related goo into the runtime module, in an
internal PromiseState class. This encapsulates the internal
state transitions and protects against misuse. It also allows
us to clean up the public API for the Property<T> type so that
it's entirely suitable for external usage.
* Track input and output property values distinctly. It turns
out we want to key off events differently. For example, to marshal
property values to a resource provider, we only care about the
inputs. For final property values that are used in, say, thens
or as inputs to other properties, we want the output property value.
* Be more precise about when an output is truly final, and known, or
unknown due to planning/dry-runs. Note that this does mean that
we'll encounter unknown values more frequently because, aside from
IDs and URNs, we can't say for sure that arbitrary properties will never
change post-creation. We have ideas on how to denote this; see
pulumi/pulumi-fabric#330 for more details.
This change adds the ability to perform runtime logging, including
debug logging, that wires up to the Pulumi Fabric engine in the usual
ways. Most stdout/stderr will automatically go to the right place,
but this lets us add some debug tracing in the implementation of the
runtime itself (and should come in handy in other places, like perhaps
the Pulumi Framework and even low-level end-user code).
This change implements free variable calculations and wires it up
to closure serialization. This is recursive, in the sense that
the serializer may need to call back to fetch free variables for
nested functions encountered during serialization.
The free variable calculation works by parsing the serialized
function text and walking the AST, applying the usual scoping rules
to determine what is free. In particular, it respects nested
function boundaries, and rules around var, let, and const scoping.
We are using Acorn to perform the parsing. I'd originally gone
down the path of using V8, so that we have one consistent parser
in the game, however unfortunately neither V8's parser nor its AST
is a stable API meant for 3rd parties. Unlike the exising internal
V8 dependencies, this one got very deep very quickly, and I became
nervous about maintaining all those dependencies. Furthermore,
by doing it this way, we can write the free variable logic in
JavaScript, which means one fewer C++ component to maintain.
This also includes a fairly significant amount of testing, all
of which passes! 🎉
This change contains an initial implementation of closure serialization
built atop V8, rather than our own custom runtime. This requires that
we use a Node.js dynamic C++ module, so that we can access the V8
APIs directly. No build magic is required beyond node-gyp.
For the most part, this was straight forward, except for one part: we
have to use internal V8 APIs. This is required for two reasons:
1) We need access to the function's lexical closure environment, so
that we may look up closure variables. Although there is a
tantalizingly-close v8::Object::CreationContext, its implementation
intentionally pokes through closure contexts in order to recover
the Function constructor context instead. That's not what we
want. We want the raw, unadulterated Function::context.
2) We need to control the lexical lookups of free variables so that
they can look past chained contexts, lexical contexts, withs, and
eval-style context extensions. Simply runing a v8::Script, or
simulating an eval, doesn't do the trick. Hence, we need to access
the unexported v8::internal::Context::Lookup function.
There is a third reason which is not yet implemented: free variable
calculation. I could use Esprima, or do my own scanner for free
variables, but I'd prefer to simply use the V8 parser so that we're
using the same JavaScript parser across all components. That too
is not part of the v8.h API, so we'll need to crack it open more.
To be clear, these are still exported public APIs, in proper headers
that are distributed with both Node and V8. They simply aren't part
of the "stable" v8.h surface area. As a result, I do expect that
maintaining this will be tricky, and I'd like to keep exploring how
to do this without needing the internal dependency. For instance,
although this works with node-gyp just fine, we will probably be
brittle across versions of Node/V8, when the internal APIs might be
changing. This will introduce unfortunate versioning headaches (all,
hopefully and thankfully, caught at compile-time).
The organization of packages underneath lib/ breaks the easy consumption
of submodules, a la
import {FileAsset} from "@pulumi/pulumi-fabric/asset";
We will go back to having everything hanging off the module root directory.
This change adds getX and requireX helper functions for configuration,
making it easy for packages to convert from Lumi's current weakly typed
config system, where everything is a string, into the internal JavaScript
representation, which is often a boolean, number, or complex array/object.
This makes a few tweaks to get the integration tests passing:
* Add `runtime: nodejs` to the minimal example's `Lumi.yaml` file.
* Remove usage of `@lumi/lumirt { printf }` and just use `console.log`.
* Remove calls to `lumijs` in the integration test framework and
the minimal example's package.json. Instead, we just run
`yarn run build`, which itself internally just invokes `tsc`.
* Add package validation logic and eliminate the pkg/compiler/metadata
library, in favor of the simpler code in pkg/engine.
* Simplify the Node.js langhost plugin CLI, and simply take an
argument rather than requiring required and optional --flags.
* Use a default path of "." if the program path isn't provided. This
is a legal scenario if you've passed a pwd and just want to load
the package's default module ("./index.js" or whatever main says).
* Add an executable script, lumi-langhost-nodejs, that fires up the
`bin/cmd/langhost/index.js` file to serve the Node.js language plugin.
This is the initial step towards redefining Lumi as a library that runs
atop vanilla Node.js/V8, rather than as its own runtime.
This change is woefully incomplete but this includes some of the more
stable pieces of my current work-in-progress.
The new structure is that within the sdk/ directory we will have a client
library per language. This client library contains the object model for
Lumi (resources, properties, assets, config, etc), in addition to the
"language runtime host" components required to interoperate with the
Lumi resource monitor. This resource monitor is effectively what we call
"Lumi" today, in that it's the thing orchestrating plans and deployments.
Inside the sdk/ directory, you will find nodejs/, the Node.js client
library, alongside proto/, the definitions for RPC interop between the
different pieces of the system. This includes existing RPC definitions
for resource providers, etc., in addition to the new ones for hosting
different language runtimes from within Lumi.
These new interfaces are surprisingly simple. There is effectively a
bidirectional RPC channel between the Lumi resource monitor, represented
by the lumirpc.ResourceMonitor interface, and each language runtime,
represented by the lumirpc.LanguageRuntime interface.
The overall orchestration goes as follows:
1) Lumi decides it needs to run a program written in language X, so
it dynamically loads the language runtime plugin for language X.
2) Lumi passes that runtime a loopback address to its ResourceMonitor
service, while language X will publish a connection back to its
LanguageRuntime service, which Lumi will talk to.
3) Lumi then invokes LanguageRuntime.Run, passing information like
the desired working directory, program name, arguments, and optional
configuration variables to make available to the program.
4) The language X runtime receives this, unpacks it and sets up the
necessary context, and then invokes the program. The program then
calls into Lumi object model abstractions that internally communicate
back to Lumi using the ResourceMonitor interface.
5) The key here is ResourceMonitor.NewResource, which Lumi uses to
serialize state about newly allocated resources. Lumi receives these
and registers them as part of the plan, doing the usual diffing, etc.,
to decide how to proceed. This interface is perhaps one of the
most subtle parts of the new design, as it necessitates the use of
promises internally to allow parallel evaluation of the resource plan,
letting dataflow determine the available concurrency.
6) The program exits, and Lumi continues on its merry way. If the program
fails, the RunResponse will include information about the failure.
Due to (5), all properties on resources are now instances of a new
Property<T> type. A Property<T> is just a thin wrapper over a T, but it
encodes the special properties of Lumi resource properties. Namely, it
is possible to create one out of a T, other Property<T>, Promise<T>, or
to freshly allocate one. In all cases, the Property<T> does not "settle"
until its final state is known. This cannot occur before the deployment
actually completes, and so in general it's not safe to depend on concrete
resolutions of values (unlike ordinary Promise<T>s which are usually
expected to resolve). As a result, all derived computations are meant to
use the `then` function (as in `someValue.then(v => v+x)`).
Although this change includes tests that may be run in isolation to test
the various RPC interactions, we are nowhere near finished. The remaining
work primarily boils down to three things:
1) Wiring all of this up to the Lumi code.
2) Fixing the handful of known loose ends required to make this work,
primarily around the serialization of properties (waiting on
unresolved ones, serializing assets properly, etc).
3) Implementing lambda closure serialization as a native extension.
This ongoing work is part of pulumi/pulumi-fabric#311.
