As a small developer quality of life improvement, I found many individual unit tests that take longer than around a second to complete, and set them to skip when run under `go test -short`.
On my machine, the wall timings for tests (with `go test -count=1 ./...` and optionally `-short` and `-race`) are roughly:
- Long tests, no race detector: about 1m42s
- Short tests, no race detector: about 17s
- Long tests, race detector enabled: about 2m1s
- Short tests, race detector enabled: about 28s
This PR is split into many commits each touching a single package, with commit messages detailing the approximate timing change per package.
shout out to @joeabbey for the inspiration. This makes the lazy
functions internal by default to prevent potential misuse by external
callers.
Should backport cleanly into 0.36 and I'll handle a messy merge into 0.35
* rpc: rework timeouts to be per-method instead of global
Prior to this change, we set a 10-second global timeout for all RPC methods
using the net/http Server type's WriteTimeout. This meant that any request
whose handler did not return within that period would simply drop the
connection to the client.
This timeout is too short for a default, as evidenced by issues like [1] and
[2]. In addition, the mode of failure on the client side is confusing; it
shows up as a dropped connection (EOF) rather than a meaningful error from the
service. More importantly, various methods have diffent constraints: Some
should be able to return quickly, others may need to adjust based on the
application workload.
This is a first step toward supporting configurable timeouts. This change:
- Removes the server-wide default global timeout, and instead:
- Wires up a default context timeout for all RPC handlers.
- Increases the default timeout from 10s to 60s.
- Adds a hook to override this per-method as needed.
This does NOT expose the timeouts in the configuration file (yet).
[1] https://github.com/osmosis-labs/osmosis/issues/1391
[2] https://github.com/tendermint/tendermint/issues/8465
In #8397 I tried to remove all the cases where we needed to keep track of the
target type of parameters for JSON encoding, but there is one case still left:
When decoding parameters from URL query terms, there is no way to tell whether
or not we need base64 encoding without knowing whether the underlying type of
the target is string or []byte.
To fix this, keep track of parameters that are []byte valued when RPCFunc is
compiling its argument map, and use that when parsing URL query terms. Update
the tests accordingly.
Pass all parameters from JSON-RPC requests to their corresponding handlers
using struct types instead of positional parameters. This allows us to control
encoding of arguments using only the standard library, and to eliminate the
remaining special-purpose JSON encoding hooks in the server.
To support existing use, the server still allows arguments to be encoded in
JSON as either an array or an object.
Related changes:
- Rework the RPCFunc constructor to reduce reflection during RPC call service.
- Add request parameter wrappers for each RPC service method.
- Update the RPC Environment methods to use these types.
- Update the interfaces and shims derived from Environment to the new
signatures.
- Update and extend test cases.
I was digging around over here, and thought it'd be good to
cleanup/standardize the line formating on a few of these methods. Also
found a few cases where we could use contexts better so did a little
bit of cleanup there too!
* testing: reduce usage of the MustDefualtLogger constructor
* Apply suggestions from code review
Co-authored-by: M. J. Fromberger <michael.j.fromberger@gmail.com>
* cleanup tests
Co-authored-by: M. J. Fromberger <michael.j.fromberger@gmail.com>
We have this one solitary metric from the go-metrics package. In principle
this statistic could be useful, but the way we have it hooked up, nothing can
observe the value: We don't export it, we don't log it, and it does not auto
publish anywhere.
Given that this state of affairs has not changed since the metric was first
added in 2017 (c08618f), I think we can safely discard it. No one is now or has
ever gotten any data out of this metric.
Extend the decoding rules for URL query parameters so that if the target type
implements encoding.TextUnmarshaler, the decoder will use it.
This is a non-breaking change.
* rpc/client: remove the placeholder RunState type.
I added the RunState type in #6971 to disconnect clients from the service
plumbing, which they do not need. Now that we have more complete context
plumbing, the lifecycle of a client no longer depends on this type: It serves
as a carrier for a logger, and a Boolean flag for "running" status, neither of
which is used outside of tests.
Logging in particular is defaulted to a no-op logger in all production use.
Arguably we could just remove the logging calls, since they are never invoked
except in tests. To defer the question of whether we should do that or make the
logging go somewhere more productive, I've preserved the existing use here.
Remove use of the IsRunning method that was provided by the RunState, and use
the Start method and context to govern client lifecycle.
Remove the one test that exercised "unstarted" clients. I would like to remove
that method entirely, but that will require updating the constructors for all
the client types to plumb a context and possibly other options. I have deferred
that for now.
* rpc: fix layout of endpoint list
The output of the default endpoint-list query was not correctly segregating
methods with and without arguments. Fix this, and also clean up the output to
be easier to read (both in code and in generated source).
Fixes#3618.
* rpc: simplify the handling of JSON-RPC request and response IDs
Replace the ID wrapper interface with plain JSON. Internally, the client
libraries use only integer IDs, and the server does not care about the ID
structure apart from checking its validity.
Basic structure of this change:
- Remove the jsonrpcid interface and its helpers.
- Unexport the ID field of request and response.
- Add helpers for constructing requests and responses.
- Fix up usage and tests.
Update the interface of the batch decoder to match the type signature of the
single-response case. The caller provides the outputs, so there is no need to
return them as well.
No functional changes.
Responses are constructed from requests using MakeResponse, MakeError, and
MakeErrorf. This ensures the response is always paired with the correct ID,
makes cases where there is no ID more explicit at the usage site, and
consolidates the handling of error introspection across transports.
The logic for unpacking errors and assigning JSON-RPC response types was
previously duplicated in three places. Consolidate it in the types package for
the RPC subsystem.
* update test cases
Instead of having the exported Request and Response type expose the version
directly, delegate version injection to an unexported shim. We already had the
shim; this just boosts it to the top level and does a bit more checking.
* Check JSON-RPC version marker.
These are only ever used with the defaults, except in our own tests. A search
of cs.github.com shows no other callers.
The use in the test was solely to bug out the go-metrics package so its
goroutines don't trigger the leak checker. Use the package's own flag for that
purpose instead. Note that calling "Stop" on the metric helps, but is not
sufficient -- the Stop does not wait for its goroutine to exit.
The main change here is to use encoding/json to encode and decode RPC
parameters, rather than the custom tmjson package. This includes:
- Update the HTTP POST handler parameter handling.
- Add field tags to 64-bit integer types to get string encoding (to match amino/tmjson).
- Add marshalers to struct types that mention interfaces.
- Inject wrappers to decode interface arguments in RPC handlers.
The parameters for RPC GET requests are parsed from query arguments in the
request URL. Rework this code to remove the need for tmjson. The structure of a
call still requires reflection, and still works the same way as before, but the
code structure has been simplified and cleaned up a bit.
Points of note:
- Consolidate handling of pointer types, so we only need to dereference once.
- Reduce the number of allocations of reflective types.
- Report errors for unsupported types rather than returning untyped nil.
Update the tests as well. There was one test case that checked for an error on
a behaviour the OpenAPI docs explicitly demonstrates as supported, so I fixed
that test case, and also added some new ones for cases that weren't checked.
Related:
* Update e2e base Go image to 1.17 (to match config).
Require that RPC functions take a context as their first argument, and return
an error as either their only result, or the second of two results.
This does not change how functions are dispatched, but will make it a little
easier to make more invasive changes in the near future.
Instead of taking a comma-separated string of parameter names, take each
parameter name as a separate argument. Now that we no longer have an extra flag
for caching, this fits nicely into a variadic trailer.
* Update all usage of NewRPCFunc and NewWSRPCFunc.
Rather than installing two separate panic handlers, defer the bookkeeping
separately from recovery, and lift the delegated handler call out to the top
level of the wrapper.
Also: Regularize the server middleware wrappers.
Add writeRPCResponse and writeHTTPResponse helpers, that handle the way RPC
responses are written to HTTP replies. These replace the exported helpers.
Visible effects:
- JSON results are now marshaled without indentation.
- HTTP status codes are now normalized.
- Cache control headers are no longer set.
Details:
- When writing a response to a URL (GET) request, do not marshal the whole
JSON-RPC object into the body, only encode the result or the error object.
This is a user-visible change.
- Do not change the HTTP status code for RPC errors. The RPC error already
reports what went wrong, the HTTP status should only report problems with the
HTTP transaction itself. This is a user-visible change.
- Encode JSON without indentation in POST response bodies. This is mainly cosmetic
but saves quite a bit of response data. Indent is still applied to GET responses to make
life easier for code examples.
- Remove an obsolete TODO about reporting an HTTP error on websocket upgrade.
Nothing needed to change; the upgrader already reports an error.
- Report an HTTP error when starting the server loop fails.
- Improve logging for encoding errors.
- Log less aggressively.
We should not set cache-control headers on RPC responses. HTTP caching
interacts poorly with resources that are expected to change frequently, or
whose rate of change is unpredictable.
More subtly, all calls to the POST endpoint use the same URL, which means a
cacheable response from one call may actually "hide" an uncacheable response
from a subsequent one. This is less of a problem for the GET endpoints, but
that means the behaviour of RPCs varies depending on which HTTP method your
client happens to use. Websocket requests were already marked statically
uncacheable, adding yet a third combination.
To address this:
- Stop setting cache-control headers.
- Update the tests that were checking for those headers.
- Remove the flags to request cache-control.
Apart from affecting the HTTP response headers, this change does not modify the
behaviour of any of the RPC methods.
* Rename rpctypes.Context to CallInfo.
Add methods to attach and recover this value from a context.Context.
* Rework RPC method handlers to accept "real" contexts.
- Replace *rpctypes.Context arguments with context.Context.
- Update usage of RPC context fields to use CallInfo.
No functional changes.
- Pull out a some helper code to simplify the control flow within the body of
the HTTP request handler.
- Front-load the URL path check so it does not get repeated for each request.
* Update Caller interface and its documentation.
* Rework MapToRequest as ParamsToRequest.
The old interface returned the result as well as populating it. Nothing was
using this, so drop the duplicated value from the return signature. Clarify the
documentation on the Caller type.
Rework the MapToRequest helper to take an arbitrary value instead of only a
map. This is groundwork for getting rid of the custom marshaling code. For now,
however, the implementation preserves the existing behaviour for the map, until
we can replace those.
These two interfaces are identical, and besides HTTPClient being confusingly
named, all but one location uses Caller. Update that one location, and drop the
redundant interface.
Apart from the tests for the websocket client, positional parameters are not
used by RPC clients. The server supports both arrays and objects, but the
client only needs to provide one or the other.
The JSON-RPC endpoint accepts requests via URL (GET) and JSON (POST). There is
no real point in having client libraries for both modes.
A search of the SDK and on GitHub suggests that most usage is via the JSON
client (via the New constructor) or websocket (NewWS), and the only uses I
found of the NewURI client constructor are in copies of our own test code.
This does not change the functionalitiy of the server, so curl and other
URL-based clients in other languages will still function as before.
This continues the push of plumbing contexts through tendermint. I
attempted to find all goroutines in the production code (non-test) and
made sure that these threads would exit when their contexts were
canceled, and I believe this PR does that.
I think calling os.Exit at arbitrary points is _bad_ and is good to
delete. I think panics in the case of data courruption have a chance
of providing useful information.
