Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Switch from Progrock to OpenTelemetry (#6835)
* progrock -> otel
* All Progrock plumbing is now gone, though we may want to bring it back
for compatibility. Removing it was a useful exercise to find the many
places where we're relying on it.
* TUI now supports -v, -vv, -vvv (configurable verbosity). Global flags
like --debug, -v, --silent, etc. are processed anywhere in the command
string and respected.
* CLI forwards engine traces and logs to configured exporters, no need
to configure engine-side (we already need this flow for the TUI)
* "Live" spans are emitted to TUI and cloud, filtered out before sending
to a traditional (non-Live) otel stack
* Engine supports pub/sub for traces and logs, can be exposed in the
future as a GraphQL subscription
* Refactor context.Background usage to context.WithoutCancel. We usually
don't want a total reset, since that drops the span context and any
other telemetry related things (loggers etc). Go 1.21 added
context.WithoutCancel which is more precise.
* engine: don't include source in slogs. Added this prospectively and it
doesn't seem worth the noise.
* idtui: DB can record multiple traces, polish
* multi traces is mostly for dagviz, so i can run it with a single DB
* add 'passthrough' UI flag which tells the UI to ignore a span and
descend into its children
* add 'ignore' UI flag, to be used sparingly for things whose
signal:noise ratio is irredeemibly low (e.g. 'id' calls)
* make loadFooFromID calls passthrough
* make Buildkit gRPC calls passthrough
* Global Progrock rogs are theoretically replaced with
tracing.GlobalLogger, but it has yet to be integrated into anything.
* Module functions are pure after all. They're already cached
per-session, so this makes DagQL reflect that, avoiding duplicate
Buildkit work that would be deduped at the Buildkit layer. Cleans up
the telemetry since previously you'd see duplicate queries.
* TODO: ensure draining is airtight
* TODO: global logging to TUI
* TODO: batch forwarded engine spans instead of emitting them "live"
* TODO: fix dagger terminal
Signed-off-by: Alex Suraci <alex@dagger.io>
* fix log draining, again, ish
previously we would cancel all subscribers for a trace whenever a
client/derver went away. but for modules/nesting this meant the inner
call would cancel the whole trace early.
* TODO: looks like services still don't drain completely?
Signed-off-by: Alex Suraci <alex@dagger.io>
* don't set up logs if not configured
Signed-off-by: Alex Suraci <alex@dagger.io>
* respect configured level
Signed-off-by: Alex Suraci <alex@dagger.io>
* clean up shim early tracing remnants
Signed-off-by: Alex Suraci <alex@dagger.io>
* synchronously detach services on main client exit
previously service spans would be left incomplete on exit. now we'll
detach from them on shutdown, which will only stop the service if we're
the last depender on it. end result _should_ be that services are always
completed through telemetry, but I've seen maybe 2 in 50 runs still
leave it running. still troubleshooting, but without this change there
is no hope at all.
fixes #6493
Signed-off-by: Alex Suraci <alex@dagger.io>
* flush telemetry before closing server clients
Honestly not 100% confirmed, but seems right. I think the final solution
might be to get traces/logs out without going through a session in the
first place.
Signed-off-by: Alex Suraci <alex@dagger.io>
* switch from errgroup to conc for panic handling
seeing a panic in ExportSpans/UploadTraces, this should help avoid
bringing whole server down - I think - or at least give us hope.
Signed-off-by: Alex Suraci <alex@dagger.io>
* nest 'starting session' beneath 'connect'
Signed-off-by: Alex Suraci <alex@dagger.io>
* send logs out from engine to log exporter too
Signed-off-by: Alex Suraci <alex@dagger.io>
* bump midterm
Signed-off-by: Alex Suraci <alex@dagger.io>
* switch to server-side telemetry pub/sub
fetching the logs/traces over a session is really annoying with draining
because the session itself gets closed before things can be fully
flushed.
Signed-off-by: Alex Suraci <alex@dagger.io>
* show newer traces first
Signed-off-by: Alex Suraci <alex@dagger.io>
* cleanup
Signed-off-by: Alex Suraci <alex@dagger.io>
* send individual Calls over telemetry instead of IDs
More than a 10x efficiency increase. Frontend still super easy to
implement.
Test:
# in ~/src/bass
$ with-dev dagger call -m ./ --src https://github.com/vito/bass unit --packages ./pkg/cli stdout --debug &> out
$ rg measuring out | cut -d= -f2 | xargs | tr ' ' '+' | sed -e 's/0m//g' -e 's/[^0-9\+]//g' | cat -v | bc
Before:
8524838 (~8.1 MiB)
After:
727039 (~0.7 MiB)
Signed-off-by: Alex Suraci <alex@dagger.io>
* idtui Base was correct in returning bool
Signed-off-by: Alex Suraci <alex@dagger.io>
* handle case where calls haven't been seen yet
kinda hacky, but it makes sense that we need to handle this, cause
loadFooFromID or generally anything can take an ID that's never been
seen by the server before, and the loadFooFromID span will come first.
Signed-off-by: Alex Suraci <alex@dagger.io>
* idtui: add space between progress and primary output
Signed-off-by: Alex Suraci <alex@dagger.io>
* swap -vvv and -vv, -vv now breaks encapsulation
Signed-off-by: Alex Suraci <alex@dagger.io>
* cleanups
Signed-off-by: Alex Suraci <alex@dagger.io>
* tidy mage
Signed-off-by: Alex Suraci <alex@dagger.io>
* tidy
Signed-off-by: Alex Suraci <alex@dagger.io>
* loosen go.mod constraints
Signed-off-by: Alex Suraci <alex@dagger.io>
* revive labels tests
Signed-off-by: Alex Suraci <alex@dagger.io>
* fix cachemap tests
Signed-off-by: Alex Suraci <alex@dagger.io>
* nuclear option: wait for all spans to complete
Rather than closing the telemetry connection and hoping the timing works
out, we keep track of which traces have active spans and wait for that
count to reach 0.
A bit more complicated but not seeing a simpler solution really. Without
this we can't ensure that the client sees the very outermost spans
complete.
Signed-off-by: Alex Suraci <alex@dagger.io>
* pass-through all gRPC stuff
hasn't really been useful, it's available in the full trace for devs, or
we can add a verbosity level.
Signed-off-by: Alex Suraci <alex@dagger.io>
* dagviz: tweaks to support visualizing a live trace
Signed-off-by: Alex Suraci <alex@dagger.io>
* better 'docker tag' parsing
Signed-off-by: Alex Suraci <alex@dagger.io>
* fixup docker tag check
Signed-off-by: Alex Suraci <alex@dagger.io>
* pass auth headers to OTLP logs too
Signed-off-by: Alex Suraci <alex@dagger.io>
* fix stdio not making it out of gateway containers
Signed-off-by: Alex Suraci <alex@dagger.io>
* fix terminal support
Signed-off-by: Alex Suraci <alex@dagger.io>
* drain immediately when interrupted
otherwise we can get stuck waiting for child spans of a nested process
that got kill -9'd. not perfect but better than hanging on Ctrl+C which
is already an emergent situation where you're not likely that interested
in any remaining data if you already had a reason to interrupt.
in Cloud we'll clean up any orphaned spans based on keepalives anyway.
Signed-off-by: Alex Suraci <alex@dagger.io>
* fix unintentionally HTTP-ifying gRPC otlp enpoint
Signed-off-by: Alex Suraci <alex@dagger.io>
* give up retrying connection if outer ctx canceled
Signed-off-by: Alex Suraci <alex@dagger.io>
* initiate draining only when main client goes away
Signed-off-by: Alex Suraci <alex@dagger.io>
* appease linter
Signed-off-by: Alex Suraci <alex@dagger.io>
* remove unnecessary wait
we don't need to try synchronizing here now that we just generically
wait for all spans to complete
Signed-off-by: Alex Suraci <alex@dagger.io>
* fix panic if no telemetry
Signed-off-by: Alex Suraci <alex@dagger.io>
* remove debug log
Signed-off-by: Alex Suraci <alex@dagger.io>
* print final progress tree in plain mode
no substitute for live console streaming, but easier to implement for
now, and probably easier to read in CI. probably needs more work, but
might get some tests passing.
Signed-off-by: Alex Suraci <alex@dagger.io>
* fix Windows build
Signed-off-by: Alex Suraci <alex@dagger.io>
* propagate spans through dagger-in-dagger
Signed-off-by: Alex Suraci <alex@dagger.io>
* retry connecting to telemetry
Signed-off-by: Alex Suraci <alex@dagger.io>
* propagate span context through dagger run
Signed-off-by: Alex Suraci <alex@dagger.io>
* install default labels as otel resource attrs
Signed-off-by: Alex Suraci <alex@dagger.io>
* tidy
Signed-off-by: Alex Suraci <alex@dagger.io>
* remove pipeline tests
these are expected to fail now
Signed-off-by: Alex Suraci <alex@dagger.io>
* fail root span when command fails
Signed-off-by: Alex Suraci <alex@dagger.io>
* Container.import: add span for streaming image
Signed-off-by: Alex Suraci <alex@dagger.io>
* idtui: break encapsulation in case of errors
Signed-off-by: Alex Suraci <alex@dagger.io>
* fix schema-level logging not exporting
caught by TestDaggerUp/random
Signed-off-by: Alex Suraci <alex@dagger.io>
* update TestDaggerRun assertion
Signed-off-by: Alex Suraci <alex@dagger.io>
* fix test not syncing on progress completion
Signed-off-by: Alex Suraci <alex@dagger.io>
* add verbose debug log
Signed-off-by: Alex Suraci <alex@dagger.io>
* respect $DAGGER_CLOUD_URL and $DAGGER_CLOUD_TOKEN
promoting these from _EXPERIMENTAL along the way, which has already been
done for _TOKEN, don't really see a strong reason to keep the
_EXPERIMENTAL prefix, but low conviction
Signed-off-by: Alex Suraci <alex@dagger.io>
* port 'processor: support span keepalive'
originally aluzzardi/otel-in-flight@2fc011f
Signed-off-by: Alex Suraci <alex@dagger.io>
* add 'watch' command
really helps with troubleshooting hanging tests!
Signed-off-by: Alex Suraci <alex@dagger.io>
* set a reasonable window size in plain mode
otherwise the terminals resize a ton of times when a long string is
printed, absolutely tanking performance. would be nice if that were
fast, but no time for that now.
Signed-off-by: Alex Suraci <alex@dagger.io>
* manually revert container.import change
i thought this wouldn't break it, but ... ?
Signed-off-by: Alex Suraci <alex@dagger.io>
* fix race
Signed-off-by: Alex Suraci <alex@dagger.io>
* mark watch command experimental
Signed-off-by: Alex Suraci <alex@dagger.io>
* fixup lock, more logging
Signed-off-by: Alex Suraci <alex@dagger.io>
* tidy
Signed-off-by: Alex Suraci <alex@dagger.io>
* fix data race in tests
Signed-off-by: Alex Suraci <alex@dagger.io>
* fix java SDK hang once again
really not sure what's writing to stderr even with --silent but this is
just too brittle. redirect stderr to /dev/null instead.
Signed-off-by: Alex Suraci <alex@dagger.io>
* retire dagger.io/ui.primary, use root span instead
fixes Views test; frontend must have been getting confused because there
were multiple "primary" spans
Signed-off-by: Alex Suraci <alex@dagger.io>
* take 2: just manually mark the 'primary' span
Signed-off-by: Alex Suraci <alex@dagger.io>
* merge tracing and telemetry packages
Signed-off-by: Alex Suraci <alex@dagger.io>
* cleanups
Signed-off-by: Alex Suraci <alex@dagger.io>
* roll back sync detach change
this was no longer needed with the change to wait for spans to finish,
not worth the review-time distraction
Signed-off-by: Alex Suraci <alex@dagger.io>
* cleanups
Signed-off-by: Alex Suraci <alex@dagger.io>
* update comment
Signed-off-by: Alex Suraci <alex@dagger.io>
* remove dead code
Signed-off-by: Alex Suraci <alex@dagger.io>
* default primary span to root span
Signed-off-by: Alex Suraci <alex@dagger.io>
* remove unused module arg
Signed-off-by: Alex Suraci <alex@dagger.io>
* send engine traces/logs to cloud
Signed-off-by: Alex Suraci <alex@dagger.io>
* implement sub metrics pub/sub
Some clients presume this service is supported by the OTLP endpoint. So
we can just have a stub implementation for now.
Signed-off-by: Alex Suraci <alex@dagger.io>
* sdk/go runtime: implement otel propagation
TODO: set up otel for you
Signed-off-by: Alex Suraci <alex@dagger.io>
* tidy
Signed-off-by: Alex Suraci <alex@dagger.io>
* add scary comment
Signed-off-by: Alex Suraci <alex@dagger.io>
* batch events that are sent from the engine
Previously we were just sending each individual update to the configured
exporters, which was very expensive and would even slow down the TUI.
When I originally tried to send it to span processors, nothing would be
sent out; turns out that was because the transform.Spans call we were
using didn't set the `Sampled` trace flag.
Now we forward engine traces and logs to all configured processors,
so their individual batching settings should be respected.
Signed-off-by: Alex Suraci <alex@dagger.io>
* fix spans being deduped within single batch
* fix detection for in flight spans; we need to check EndTime <
StartTime since sometimes we end up with a 1754 timestamp
* when a span is already present in a batch, update it in-place rather
than dropping it on the floor
Signed-off-by: Alex Suraci <alex@dagger.io>
* Add Python support
Signed-off-by: Helder Correia <174525+helderco@users.noreply.github.com>
* shim: proxy otel to 127.0.0.1:0
more universally compatible than unix://
Signed-off-by: Alex Suraci <alex@dagger.io>
* remove unnecesssary fn
Signed-off-by: Alex Suraci <alex@dagger.io>
* attributes: add passthrough, bikeshed + document
also start cleaning up "tasks" cruft nonsense, these can just be plain
old attributes on a single span i think
Signed-off-by: Alex Suraci <alex@dagger.io>
* fix janky flag parsing
parse global flags in two passes, ensuring the same flags are installed
in both cases, and capturing the values before installing them into the
real flag set, since that clobbers the values
Signed-off-by: Alex Suraci <alex@dagger.io>
* discard Buildkit progress
...just in case it gets buffered in memory forever otherwise
Signed-off-by: Alex Suraci <alex@dagger.io>
* sdk/go: somewhat gross support for opentelemetry
had to copy-paste a lot of the telemetry code into sdk/go/. would love
to just move everything there so it can be shared between the shim, the
Go runtime, and the engine, however it is currently a huge PITA to share
code between all three, because of the way codegen works. saving that
for another day. maybe tomorrow.
Signed-off-by: Alex Suraci <alex@dagger.io>
* send logs to function call span, not exec /runtime
Signed-off-by: Alex Suraci <alex@dagger.io>
* tui: respect dagger.io/ui.mask
no more exec /runtime!
Signed-off-by: Alex Suraci <alex@dagger.io>
* silence linter
worth refactoring, but not now™
Signed-off-by: Alex Suraci <alex@dagger.io>
* ignore --help when parsing global flags
Signed-off-by: Alex Suraci <alex@dagger.io>
* Pin python requirements
Signed-off-by: Helder Correia <174525+helderco@users.noreply.github.com>
* revert Python SDK changes for now
looks like there's more to figure out with module dependencies? either
way, don't want this to block the current PR, they can be re-introduced
in another PR like the other SDKs
Revert "Pin python requirements"
This reverts commit b40c411.
Revert "Add Python support"
This reverts commit 08aa92c.
Signed-off-by: Alex Suraci <alex@dagger.io>
* fix race conditions in python SDK runtime
Signed-off-by: Alex Suraci <alex@dagger.io>
---------
Signed-off-by: Alex Suraci <alex@dagger.io>
Signed-off-by: Helder Correia <174525+helderco@users.noreply.github.com>
Co-authored-by: Helder Correia <174525+helderco@users.noreply.github.com>- Loading branch information
