As of this change, the engine will run all Configure calls in parallel.
This improves startup performance, since otherwise, we would block
waiting for all plugins to be configured before proceeding to run a
program. Emperically, this is about 1.5-2s for AWS programs, and
manifests as a delay between the purple "Previewing update of stack"
being printed, and the corresponding grey "Previewing update" message.
This is done simply by using a Goroutine for Configure, and making sure
to synchronize on all actual CRUD operations. I toyed with using double
checked locking to eliminate lock acquisitions -- something we may want
to consider as we add more fine-grained parallelism -- however, I've
kept it simple to avoid all the otherwise implied memory model woes.
I made the judgment call that GetPluginInfo may proceed before
Configure has settled. (Otherwise, we'd immediately call it and block
after loading the plugin, obviating the parallelism benefits.) I also
made the judgment call to do this in the engine, after flip flopping
several times about possibly making it a provider's own decision.
* Refactor the SnapshotManager interface
Lift snapshot management out of the engine by delegating it to the
SnapshotManager implementation in pkg/backend.
* Add a event interface for plugin loads and use that interface to record plugins in the snapshot
* Remove dead code
* Add comments to Events
* Add a number of tests for SnapshotManager
* CR feedback: use a successful bit on 'End' instead of having a separate 'Abort' API
* CR feedback
* CR feedback: register plugins one-at-a-time instead of the entire state at once
This issue arises becuase the behavior we're currently getting from Diff
for TF-based providers differs from the behavior we expect. We are
presenting the provider with the old state and new inputs. If the old
state contains output properties that differ from the new inputs, the
provider will detect a diff where we may expect no changes.
Rather than deferring to the provider for all diffs, these changes only
defer to the provider if a legacy diff was detected (i.e. there is some
difference between the old and new provider-calculated inputs).
This change implements a `pulumi refresh` command. It operates a bit
like `pulumi update`, and friends, in that it supports `--preview` and
`--diff`, along with the usual flags, and will update your checkpoint.
It works through substitution of the deploy.Source abstraction, which
generates a sequence of resource registration events. This new
deploy.RefreshSource takes in a prior checkpoint and will walk it,
refreshing the state via the associated resource providers by invoking
Read for each resource encountered, and merging the resulting state with
the prior checkpoint, to yield a new resource.Goal state. This state is
then fed through the engine in the usual ways with a few minor caveats:
namely, although the engine must generate steps for the logical
operations (permitting us to get nice summaries, progress, and diffs),
it mustn't actually carry them out because the state being imported
already reflects reality (a deleted resource has *already* been deleted,
so of course the engine need not perform the deletion). The diffing
logic also needs to know how to treat the case of refresh slightly
differently, because we are going to be diffing outputs and not inputs.
Note that support for managed stacks is not yet complete, since that
requires updates to the service to support a refresh endpoint. That
will be coming soon ...
This change removes the need to `pulumi init` when targeting the local
backend. A fair amount of the change lays the foundation that the next
set of changes to stop having `pulumi init` be used for cloud stacks
as well.
Previously, `pulumi init` logically did two things:
1. It created the bookkeeping directory for local stacks, this was
stored in `<repository-root>/.pulumi`, where `<repository-root>` was
the path to what we belived the "root" of your project was. In the
case of git repositories, this was the directory that contained your
`.git` folder.
2. It recorded repository information in
`<repository-root>/.pulumi/repository.json`. This was used by the
cloud backend when computing what project to interact with on
Pulumi.com
The new identity model will remove the need for (2), since we only
need an owner and stack name to fully qualify a stack on
pulumi.com, so it's easy enough to stop creating a folder just for
that.
However, for the local backend, we need to continue to retain some
information about stacks (e.g. checkpoints, history, etc). In
addition, we need to store our workspace settings (which today just
contains the selected stack) somehere.
For state stored by the local backend, we change the URL scheme from
`local://` to `local://<optional-root-path>`. When
`<optional-root-path>` is unset, it defaults to `$HOME`. We create our
`.pulumi` folder in that directory. This is important because stack
names now must be unique within the backend, but we have some tests
using local stacks which use fixed stack names, so each integration
test really wants its own "view" of the world.
For the workspace settings, we introduce a new `workspaces` directory
in `~/.pulumi`. In this folder we write the workspace settings file
for each project. The file name is the name of the project, combined
with the SHA1 of the path of the project file on disk, to ensure that
multiple pulumi programs with the same project name have different
workspace settings.
This does mean that moving a project's location on disk will cause the
CLI to "forget" what the selected stack was, which is unfortunate, but
not the end of the world. If this ends up being a big pain point, we
can certianly try to play games in the future (for example, if we saw
a .git folder in a parent folder, we could store data in there).
With respect to compatibility, we don't attempt to migrate older files
to their newer locations. For long lived stacks managed using the
local backend, we can provide information on where to move things
to. For all stacks (regardless of backend) we'll require the user to
`pulumi stack select` their stack again, but that seems like the
correct trade-off vs writing complicated upgrade code.
* Lift snapshot management out of the engine
This PR is a prerequisite for parallelism by addressing a major problem
that the engine has to deal with when performing parallel resource
construction: parallel mutation of the global snapshot. This PR adds
a `SnapshotManager` type that is responsible for maintaining and
persisting the current resource snapshot. It serializes all reads and
writes to the global snapshot and persists the snapshot to persistent
storage upon every write.
As a side-effect of this, the core engine no longer needs to know about
snapshot management at all; all snapshot operations can be handled as
callbacks on deployment events. This will greatly simplify the
parallelization of the core engine.
Worth noting is that the core engine will still need to be able to read
the current snapshot, since it is interested in the dependency graphs
contained within. The full implications of that are out of scope of this
PR.
Remove dead code, Steps no longer need a reference to the plan iterator that created them
Fixing various issues that arise when bringing up pulumi-aws
Line length broke the build
Code review: remove dead field, fix yaml name error
Rebase against master, provide implementation of StackPersister for cloud backend
Code review feedback: comments on MutationStatus, style in snapshot.go
Code review feedback: move SnapshotManager to pkg/backend, change engine to use an interface SnapshotManager
Code review feedback: use a channel for synchronization
Add a comment and a new test
* Maintain two checkpoints, an immutable base and a mutable delta, and
periodically merge the two to produce snapshots
* Add a lot of tests - covers all of the non-error paths of BeginMutation and End
* Fix a test resource provider
* Add a few tests, fix a few issues
* Rebase against master, fixed merge
We've seen failures in CI where DNS lookups fail which cause our
operations against the service to fail, as well as other sorts of
timeouts.
Add a set of helper methods in a new httputil package that helps us do
retries on these operations, and then update our client library to use
them when we are doing GET requests. We also provide a way for non GET
requests to be retried, and use this when updating a lease (since it
is safe to retry multiple requests in this case).
The RPC provider interface needs a way to convey back to the engine
that a resource being read no longer exists. To do this, we'll return
the ID property that was read back. If it is empty, it means the
resource is gone. If it is non-empty, we expect it to match the input.
This change skips unknown IDs during read operations. This can happen
when a read is performed using the output property of another resource
during planning. This is intentionally supported via ID being an
Input<ID> and all we need to do for this to work correctly is skip the
actual provider RPC and the runtime will propagate unknown outputs as
usual.
This change lets plugin versions to float in two ways:
1) If a `pulumi plugin install` detects a newer version is available
already, there's no need to download and install the older version.
2) If the engine attempts to load a plugin at a particular version,
if a newer version is available, it will be accepted without error.
As part of this, we permit $PATH to have the final say when determining
which version to accept. That is, it can always override the choice.
Note that I highly suspect, in the limit, that we'll want to stop doing
this for major version incompatibilities. For now, since we don't
envision any such version changes imminently, this will suffice.
This change wires up the new Read RPC method in such a manner that
Pulumi programs can invoke it. This is technically not required for
refreshing state programmatically (as in pulumi/pulumi#1081), however
it's a feature we had eons ago and have wanted since (see
pulumi/pulumi#83), and will allow us to write code like
let vm = aws.ec2.Instance.get("my-vm", "i-07043cd97bd2c9cfc");
// use any property from here on out ...
The way this works is simply by bridging the Pulumi program via its
existing RPC connection to the engine, much like Invoke and
RegisterResource RPC requests already do, and then invoking the proper
resource provider in order to read the state. Note that some resources
cannot be uniquely identified by their ID alone, and so an extra
resource state bag may be provided with just those properties required.
This came almost for free (okay, not exactly) and will come in handy as
we start gaining experience with reading live state from resources.
This commit changes two things about our resource model:
* Stop performing Pulumi Engine-side diffing of resource state.
Instead, we defer to the resource plugins themselves to determine
whether a change was made and, if so, the extent of it. This
manifests as a simple change to the Diff function; it is done in
a backwards compatible way so that we continue with legacy diffing
for existing resource provider plugins.
* Add a Read RPC method for resource providers. It simply takes a
resource's ID and URN, plus an optional bag of further qualifying
state, and it returns the current property state as read back from
the actual live environment. Note that the optional bag of state
must at least include enough additional properties for resources
wherein the ID is insufficient for the provider to perform a lookup.
It may, however, include the full bag of prior state, for instance
in the case of a refresh operation.
This is part of pulumi/pulumi#1108.
* Improve the error message arising from missing required configs for
resource providers
If the resource provider that we are speaking to is new enough, it will send
across a list of keys and their descriptions alongside an error
indicating that the provider we are configuring is missing required
config. This commit packages up the list of missing keys into an error
that can be presented nicely to the user.
* Code review feedback: renaming simplification and correcting errors in comments
* Send structured errors across RPC boundaries
This brings us closer to gRPC best practices where we send structured
errors with error codes across RPC endpoints. The new "rpcerrors"
package can wrap errors from RPC endpoints, so RPC servers can attach
some additional context as to why a request failed.
* Code review feedback:
1. Rename rpcerrors -> rpcerror, better package name
2. Rename RPCError -> Error, RPCErrorCause -> ErrorCause, names
suggested by gometalinter to improve their package-qualified names
3. Fix import organization in rpcerror.go
This change includes a bunch of refactorings I made in prep for
doing refresh (first, the command, see pulumi/pulumi#1081):
* The primary change is to change the way the engine's core update
functionality works with respect to deploy.Source. This is the
way we can plug in new sources of resource information during
planning (and, soon, diffing). The way I intend to model refresh
is by having a new kind of source, deploy.RefreshSource, which
will let us do virtually everything about an update/diff the same
way with refreshes, which avoid otherwise duplicative effort.
This includes changing the planOptions (nee deployOptions) to
take a new SourceFunc callback, which is responsible for creating
a source specific to the kind of plan being requested.
Preview, Update, and Destroy now are primarily differentiated by
the kind of deploy.Source that they return, rather than sprinkling
things like `if Destroying` throughout. This tidies up some logic
and, more importantly, gives us precisely the refresh hook we need.
* Originally, we used the deploy.NullSource for Destroy operations.
This simply returns nothing, which is how Destroy works. For some
reason, we were no longer doing this, and instead had some
`if Destroying` cases sprinkled throughout the deploy.EvalSource.
I think this is a vestige of some old way we did configuration, at
least judging by a comment, which is apparently no longer relevant.
* Move diff and diff-printing logic within the engine into its own
pkg/engine/diff.go file, to prepare for upcoming work.
* I keep noticing benign diffs anytime I regenerate protobufs. I
suspect this is because we're also on different versions. I changed
generate.sh to also dump the version into grpc_version.txt. At
least we can understand where the diffs are coming from, decide
whether to take them (i.e., a newer version), and ensure that as
a team we are monotonically increasing, and not going backwards.
* I also tidied up some tiny things I noticed while in there, like
comments, incorrect types, lint suppressions, and so on.
This change uses the prior checkpoint's deployment manifest to pre-
populate all plugins required to complete the destroy operation. This
allows for subsequent attempts to load a resource's plugin to match the
already-loaded version. This approach obviously doesn't work in a
hypothetical future world where plugins for the same resource provider
are loaded side-by-side, but we already know that.
This takes the existing `apitype.Checkpoint` type and renames it to
`apitype.CheckpointV1` locking in the shape. In addition, we introduce
a `apitype.VersionedCheckpoint` type, which holds a version number and
a json document representing a checkpoint at that version. Now, when
reading a checkpoint, the CLI can determine if it's in a format it
understands, and fail gracefully if it is not.
While the CLI understands the older checkpoint version, it always
writes the newest version format, meaning that if you manage a
fire-and-forget stack with this version of the CLI, it will be
un-readable by previous versions.
Stacks managed by Pulumi.com are not impacted by this change.
Fixes: #887
* Improve error messages output by the CLI
This fixes a couple known issues with the way that we present errors
from the Pulumi CLI:
1. Any errors from RPC endpoints were bubbling up as they were to
the top-level, which was unfortunate because they contained
RPC-specific noise that we don't want to present to the user. This
commit unwraps errors from resource providers.
2. The "catastrophic error" message often got printed twice
3. Fatal errors are often printed twice, because our CLI top-level
prints out the fatal error that it receives before exiting. A lot of
the time this error has already been printed.
4. Errors were prefixed by PU####.
* Feedback: Omit the 'catastrophic' error message and use a less verbose error message as the final error
* Code review feedback: interpretRPCError -> resourceStateAndError
* Code review feedback: deleting some commented-out code, error capitalization
* Cleanup after rebase
When a stack has secrets, we now take the secret values and construct
a regular expression which is just an alternation of all the secret
values. Then, before pushing any string data into an Event, we run the
regular expression and replace all matches with '[secret]'.
Fixes#747
As it stands, we allow plugin load requests to race. Not only does this
create a situation in which we may load and then immediately throw away
a plugin (potentially leaking its process), it also creates the
possibility of races when reading from/writing to the various plugin
caches. These changes serialize all plugin loads and cache accesses by
running all accesses for a particular host in a single goroutine.
Fixes#1020.
This helper method is only really used for testing, but we should not
allow it to create a Key who's namespace has a colon (as ParseKey
would not build something like this).
This API was introduced to aid the refactoring, but it isn't something
we want to support long term. Remove it and for a few places, push
passing config.Key around more, instead of converting to the old type
eagerly.
When serializing config.Key's we now write them as <package>:<name>
instead of <package>:config:<name>. We continue to support reading the
older format for compatability with older files.
config.Key has become a pair of namespace and name. Because the whole
world has not changed yet, there continues to be a way to convert
between a tokens.ModuleMember and config.Key, however now sometime the
conversion from tokens.ModuleMember can fail (when the module member
is not of the form `<package>:config:<name>`).
I'll be changing the structure of the representation of config.Key, so
let's write some tests first to ensure we can continue to treat
everything as JSON and YAML.
Right now, config.Key is a type alias for tokens.ModuleMember. I did a
pass over the codebase such that we use config.Key everywhere it
looked like the value did not leak to some external process (e.g a
resource provider or a langhost).
Doing this makes it a little clearer (hopefully) where code is
depending on a module member structure (e.g. <package>:config:<value>)
instead of just an opaque type.
As it stands, we only configure those providers for which configuration
is present. This can lead to surprising failure modes if those providers
are then used to create resources. These changes ensure that all
resource providers that are not configured during plan initialization
are configured upon first load.
Fixes#758.
A hold-over from a previous experiment (LumiIDL) which we don't use
anymore. If we decide to bring that back, we can easily restore these
types, but for now, let's just remove this dead code.
Most of the errors in this package are holdovers from our previous
syetem where we had our own custom compiler and evaluator and are no
longer needed. The few we still use during plan applicaton (via the
diagnostics system, which is another component from the old system
that we still use) have been promoted into the diag package. Doing so,
allows us to not have to import "github.com/pkg/errors" as "goerr" in
some parts of the engine, a nice cleaup.
Despite our good progress moving towards having an apitype package,
where our exchange types live and can be shared among the engine and
our services, there were a few major types that were still duplciated.
Resource was the biggest example -- and indeed, the apitype varirant
was missing the new Dependencies property -- but there were others,
like Manfiest, PluginInfo, etc. These too had semi-random omissions.
This change merges all of these types into the apitype package. This
not only cleans up the redundancy and missing properties, but will
"force the issue" with respect to keeping them in sync and properly
versioning the information in a backwards compatible way.
The resource/stack package still exists as a simple marshaling layer
to and from the engine's core data types.
Finally, I've made the controversial change to share the actual
Deployment data structure at the apitype layer also. This will force
us to confront differences in that data structure similarly, and will
allow us to leverage the strong typing throughout to catch issues.
Previously, we would prefer a plugin on the $PATH which is more or
less always the case for people hacking on `pulumi`. Later, when we
went to check the loaded plugin version matched the one we requested,
we fail.
Now, if we have a version, we'll first consult the local plugin
cache. If that fails, we'll fall back to the $PATH as we used to.
When we are loading a plugin without a version, we continue to use the
one on the $PATH (without testing the cache) on the assumption it is
newer.
In addition, we've turned the "plugin versions are mis-matched" from
an error into a warning. We expect that we'll only ever see this
warning when something strange is going on (since in the normal case,
we'll have found the exact version in the cache) but having it not
hard fail does help in development cases.
Fixes#977
