As discovered in #7224, Json RPC protocol was added to the asynchronous resource initialization stage, as part of #6306, but was not in fact initialized at that point.
Instead it was initialized when the server was started to be able to serve correctly the initialization messages. A classic Catch-22. It was really hard to discover this just by looking at the code, but the profiling clearly showed where the time was spent.
This change splits Language Server's protocol into two:
- the first one accepts `heartbeat/init` and `session/initProtocolConnection`
- the second one enriches it with the full set of supported messages
This shifts the initialization from blocking for 0.5 sec to only ~30ms, and performing the second stage asynchronously.
Closes#7224.
# Important Notes
Before the change (blocking server startup):
After the change (1st stage):
After the change (2nd, asynchronous initialization, stage):
Package's config information, once loaded, never changed. While there is typically no need for it, this was problematic when the config became out-of-sync with the filesystem, like in the case of project rename action.
In rename, the config's properties would be updated in the FS, but that would never be reflected in module's package. Therefore further compilations would continue to ask for the old namespace.
Most of the changes are cosmetic (s/`.config`/`.getConfig()`) except for the new `reloadConfig` method on `Package` that is being called in `RenameProjectCmd` handler.
Closes#7062.
# Important Notes
The reported `ExecutionFailed` error should have been mostly fixed already via #7143. This change makes sure that all the related warnings are gone as well and the compiler uses the updated namespace.
`executionFailed` instead is sent when an evaulation finishes with a a critical failure or a non-critical error.
The PR tries to miniminally modify the change in the messages exchange so as to avoid a major redesign at this point.
Closes#7002.
# Important Notes
Unblocks IDE which will need to modify to this new setup.
Fixes#6955 by:
- using `visualisationModule` to specify the module where the visualization is to be used
- referring to method in `Meta.get_annotation` with `.method_name` - e.g. unresolved symbol notation
- evaluating arguments to `Meta.get_annotation` in the context of the user module (which can access the extension functions)
Request Timeouts started plaguing IDE due to numerous `executionContext/***Visualization` requests. While caused by a bug they revealed a bigger problem in the Language Server when serving large amounts of requests:
1) Long and short lived jobs are fighting for various locks. Lock contention leads to some jobs waiting for a longer than desired leading to unexpected request timeouts. Increasing timeout value is just delaying the problem.
2) Requests coming from IDE are served almost instantly and handled by various commands. Commands can issue further jobs that serve request. We apparently have and always had a single-thread thread pool for serving such jobs, leading to immediate thread starvation.
Both reasons increase the chances of Request Timeouts when dealing with a large number of requests. For 2) I noticed that while we used to set the `enso-runtime-server.jobParallelism` option descriptor key to some machine-dependent value (most likely > 1), the value set would **only** be available for instrumentation. `JobExecutionEngine` where it is actually used would always get the default, i.e. a single-threaded ThreadPool. This means that this option descriptor was simply misused since its introduction. Moved that option to runtime options so that it can be set and retrieved during normal operation.
Adding parallelism intensified problem 1), because now we could execute multiple jobs and they would compete for resources. It also revealed a scenario for a yet another deadlock scenario, due to invalid order of lock acquisition. See `ExecuteJob` vs `UpsertVisualisationJob` order for details.
Still, a number of requests would continue to randomly timeout due to lock contention. It became apparent that
`Attach/Modify/Detach-VisualisationCmd` should not wait until a triggered `UpsertVisualisationJob` sends a response to the client; long and short lived jobs will always compete for resources and we cannot guarantee that they will not timeout that way. That is why the response is sent immediately from the command handler and not from the job executed after it.
This brings another problematic scenario:
1. `AttachVisualisationCmd` is executed, response sent to the client, `UpsertVisualisationJob` scheduled.
2. In the meantime `ModifyVisualisationCmd` comes and fails; command cannot find the visualization that will only be added by `UpsertVisualisationJob`, which might have not yet been scheduled to run.
Remedied that by checking visualisation-related jobs that are still in progress. It also allowed for cancelling jobs which results wouldn't be used anyway (`ModifyVisualisationCmd` sends its own `UpsertVisualisationJob`). This is not a theoretical scenario, it happened frequently on IDE startup.
This change does not fully solve the rather problematic setup of numerous locks, which are requested by short and long lived jobs. A better design should still be investigated. But it significantly reduces the chances of Request Timeouts which IDE had to deal with.
With this change I haven't been able to experience Request Timeouts for relatively modest projects anymore.
I added the possibility of logging wait times for locks to better investigate further problems.
Closes#7005
close#6936
Changelog:
- add: new suggestion type Getter that is not exposed to the api
- update: do not return suggestion of type getter when doing a global search (without specifying self types)
Private suggestions and modules mentioned in the issue will be filtered out after we finish the work on the new (refined) exports algorithm.
# Important Notes
close#6800
Update the `executionContext/expressionUpdates` notification and send the list of not applied arguments in addition to the method pointer.
# Important Notes
IDE is updated to support the new API.
The change adds an additional field to `ExpressionUpdates` messages sent by `ProgramExecutionSupport` to indicate if the type of value (or its method pointer) has changed and therefore would potentially require a suggestions' update.
Prior to #3729 that check was done during the instrumentation. However we still want to continue to support "pending expression" functionality therefore `SuggestionsHandler` will use the additional information to filter only the required expression updates.
Most of the changes are related to adapting our tests to the new field.
Closes#6706.
# Important Notes
The associated project now loads and navigates smoothly.
Also attaching a screenshot from the project that illustrates that pending functionality continues to work:
[Kazam_screencast_00006.webm](https://github.com/enso-org/enso/assets/292128/35918841-f84f-4e1c-b1b0-40e45d97e111)
Artifically limiting the number of reported warnings to 100. Also added benchmarks with random Ints to investigate perf issues when dealing with warnings (future task).
Ideally we would have a custom set-like collection that allows us internally to specify a maximal number of elements. But `EnsoHashMap` (and potentially `EnsoSet`) are still WIP when it comes to being PE-friendly.
The change also allows for checking if the limit for the number of reported warnings has been reached. It will visualize by adding an additional "Warnings limit reached." to the visualization.
The limit is configurable via `--warnings-limit` parameter to `run`.
Closes#6283.
Dead Letter logging is occasionally flooding our logs which is confusing to users reporting bugs. Left the possibility of a single report so that we know that something is happening.
During initialization JGit may attempt to resolve hostname. On some systems this can take more than desired triggering timeouts. This change does two things:
- sets the default committer for changes, lack of which probably triggered the check
- sets the default hostname to `localhost` (we don't care), in case something else in JGit still wants to resolve hostname
Closes#6447.
# Important Notes
I wasn't able to reproduce this so relying on @mwu-tow since apparently he can repro it reliably.
close#6306
Changelog:
- add: `AsyncResourceInitialization` component
- update: initialize language server resources asynchronously
# Important Notes
Speeds up `session/InitProtocolConnection` request by ~300ms (~10%) 3051ms (before) vs 2740ms (after) on my machine.
close#6232
Changelog:
- remove: `SqlVersionsRepo`
- update: `SuggestionsDatabaseModuleUpdateNotification` message removing the version
- update: cleanup versions repo usages in the language server
close#6080
Changelog
- add: implement `SuggestionsRepo.insertAll` as a batch SQL insert
- update: `search/getSuggestionsDatabase` returns empty suggestions. Currently, the method is only used at startup and returns the empty response anyway because the libs are not loaded at that point.
- update: serialize only global (defined in the module scope) suggestions during the distribution building. There's no sense in storing the local library suggestions.
- update: sqlite dependency
- remove: unused methods from `SuggestionsRepo`
- remove: Arguments table
# Important Notes
Speeds up libraries loading by ~1 second.
close#6139close#6137
When the project is renamed, the engine cleans up affected modules and initiates modules re-indexing to fill the suggestions database with new records. This way it reduces the amount of information stored in the suggestions database and helps implement #6080 optimization.
Changelog:
- remove: rename features from the suggestions database
- update: rename command to initiate modules cleanup and project re-execution
- fix: #6137
It is sometimes impossible to figure out the real reason for invalid text edit request. Added a bit of context to failures to narrow down the cause of the failure.
# Important Notes
Should help with diagnosing issues like #6099.
close#5874
Changelog:
- add: `isStatic` parameter to `search/completion` request to search by the `static` suggestion attribute
- update: search non-static suggestions when opening component browser
# Important Notes
Component browser doesn't show `Table.new` and `Table.from_rows` suggestions when a `Table` node is selected.
close#5881
Changelog
- add: ProtocolFactory object that initializes and returns the protocol object
- update: Add protocol initialization to the initialization component
close#5911
In interactive mode, perform writing IR caches in the background jobs queue. Background jobs execution is delayed until the first execution is complete.
The `logAvailableComponentsForDebugging` will check and install all necessary components of GraalVM for every mentioned version. While not harmful, it adds up to startup time.
Additionally added an option in language server startup to skip installation of GraalVM components. The latter is already performed by project-manager when opening the project and it is unnecessary to do it twice. Due to LS' architecture this configuration has to be passed around via multiple configs.
Finally, skipped the attempt to install Python component on Windows - this is not supported by GraalVM atm.
Closes#5749.
# Important Notes
The impact of this problem could be really felt the more versions of Enso and GraalVM one had since it would go through all of them.
Implement new Enso documentation parser; remove old Scala Enso parser.
Performance: Total time parsing documentation is now ~2ms.
# Important Notes
- Doc parsing is now done only in the frontend.
- Some engine tests had never been switched to the new parser. We should investigate tests that don't pass after the switch: #5894.
- The option to run the old searcher has been removed, as it is obsolete and was already broken before this (see #5909).
- Some interfaces used only by the old searcher have been removed.
close#5892
Changelog:
add: feature to delay background jobs execution
add: start background jobs when program finishes
add: start background jobs on `search/completion` request
close#5070
Changelog:
- Include the original exception to log expressions
- Enable logging of Akka Actors' lifecycle events on debug logging level
- Decrease the severity of interruption log messages because interruptions are part of the workflow. The computation can be interrupted at any time, and still be recomputed after. Warnings are just misleading in this case.
This change adds serialization and deserialization of library bindings.
In order to be functional, one needs to first generate IR and
serialize bindings using `--compiled <path-to-library>` command. The bindings
will be stored under the library with `.bindings` suffix.
Bindings are being generated during `buildEngineDistribution` task, thus not
requiring any extra steps.
When resolving import/exports the compiler will first try to load
module's bindings from cache. If successful, it will not schedule its
imports/exports for immediate compilation, as we always did, but use the
bindings info to infer the dependent modules.
The current change does not make any optimizations when it comes to
compiling the modules, yet. It only delays the actual
compilation/loading IR from cache so that it can be done in bulk.
Further optimizations will come from this opportunity such as parallel
loading of caches or lazily inferring only the necessary modules.
Part of https://github.com/enso-org/enso/issues/5568 work.
Implements the #5643 idea. As soon as `MainModule` creates `Context` for GraalVM execution, it schedules a background task to initialize JavaScript. The initialization finishes sooner than Enso compiler is ready to work, saving time when it is actually needed.
# Important Notes
Only modifies boot sequence of `MainModule` (used in the IDE) and `VerifyJavaScriptIsAvailableTest` (to verify the _"context passing logic"_ works OK between threads). Regular CLI execution remains unchanged for now assuming batch execution may not need JavaScript in all the cases and if it does the initialization speed isn't that critical.
When the `project.yaml` specifies an unknown edition, the project is started with the fallback edition (current engine version). It works this way:
- IDE sends project/open request to project manager (specifying only which project to open)
- project manager resolves the project edition (based on `project.yaml`) or selects the fallback if the requested edition cannot be loaded
- project manager starts the language server specifying the edition with the `--server-edition` command line parameter
- language server specifies the correct edition when creates the runtime context
Closes#5036
Move the logic that looks up method pointers from the language server to IDE. This way we can keep the suggestion updates and expression updates async, otherwise it will hurt the initial startup time of LS.
Fixes the issue when some expression updates does not contain the method pointer.
Follow up on
16ba57d465 which marked the tests as flaky.
This test goes back to the original implementation where JGit did all its book-keeping synchronously. When we switched to asynchronous operation, cleaning up the test directory would sometimes not succeed because there would be a race-condition with the ongoing work of JGit.
# Important Notes
In the current setup tagging a test as flaky is only valid for Windows. The other platforms ignore the flag. So we were still getting random failures.
Start `project-manager` with following options to provide first 20s of the startup sequence:
```
$ project-manager --profiling-events-log-path=start.log --profiling-path=start.npss --profiling-time=20
```
once the `start.log` and `start.npss` files are generated (next to each other), open them in GraalVM's VisualVM:
```
$ graalvm/bin/jvisualvm --openfile start.npss
```
analyze.
Implementation of https://www.pivotaltracker.com/story/show/184012743https://user-images.githubusercontent.com/919491/214082311-cf49e43c-1d1f-4654-903c-a4224cd954d8.mp4
This is also a step towards more general widget support. The widget metadata is queried using `Meta.get_annotation` method through a dedicated visualization. For now only `Single_Choice` case is handled, and always all suggestions are is returned.
# Important Notes
There are limitations as to which node segments receive a widget. Only chain method calls are supported now (`thing.method` syntax), and only outside of lambda scope. Widgets in lambdas will require support for visualisations of lambda subexpressions, which is currently missing in the engine. The IDE technically tries to place the widgets there, but the data never arrives. It should work once the engine support is added.
This PR includes a mock for `Meta.get_annotation` call that only supports `Table.at` method. Real implementation is a separate task that is already in progress.
JGit ops generally run fast (as in a few milliseconds) except for the first commit. The initialization + first commit was taking at least 3.5 seconds constistenly, but only in the first test case. Now, this led to frequent timeouts down the chain when the request was expected to finish fast.
The bookkeeping involved some timestamping and other expensive calls in order to calculate clock drift. The default appears to be to run it in a blocking mode, hence adding at least 3 seconds to the first command call.
Setting the job to run in the background makes the cost of the repo initialization acceptable (~300 milliseconds on a cold JVM). The other commands are unaffected and take < 10 milliseconds.
# Important Notes
Added a test to ensure that we don't introduce the regression. Marked it as potentially flaky because it uses timestamps and it is therefore prone to random system hiccups.
There is no need for JGit to try to save file attributes to its config in `<home>/jgit/config`. This change ensures that JGit operates on a stub file, without polluting users' configs, similarly to user's git config.
Additionally, noticed that all FS operations during `init` are rather slow and saw at least on one occassion when the handler timed out because of that (https://www.pivotaltracker.com/story/show/184359934). This change provides an init-specific timeout for `vcs/init`.
# Important Notes
This change should reduce chances of failures during VCS initialization.
When a large long would be passed to a host call expecting a double, it would crash with a
```
Cannot convert '<some long>'(language: Java, type: java.lang.Long) to Java type 'double': Invalid or lossy primitive coercion
```
That is unlikely to be expected by users. It also came up in the Statistics examples during Sum. One could workaround it by forcing the conversion manually with `.to_decimal` but it is not a permanent solution.
Instead this change adds a custom type mapping from Long to Double that will do it behind the scenes with no user interaction. The mapping kicks in only for really large longs.
# Important Notes
Note that the _safe_ range is hardcoded in Truffle and it is not accessible in enso packages. Therefore a simple c&p for that max safe long value was necessary.
