Automatically add Trace Spans to Go methods and functions

This tool uses standard Go library to modify AST with instrumentation. You can add new instrumentations by defining your own Instrumenter and invoking Processor like it is done in main.

• No dependencies
• 500 LOC
• OpenTelemetry

go install github.com/nikolaydubina/go-instrument@latest

find . -name "*.go" | xargs -I{} go-instrument -app my-service -w -filename {}

Functions and methods with ctx context.Context in arguments

func (s Cat) Name(ctx context.Context) (name string, err error) {
  ...

will be instrumented with span

func (s Cat) Name(ctx context.Context) (name string, err error) {
	ctx, span := otel.Trace("my-service").Start(ctx, "Cat.Name")
	defer span.End()
	defer func() {
		if err != nil {
			span.SetStatus(codes.Error, "error")
			span.RecordError(err)
		}
	}()
  ...

Example HTTP server go-instrument-example as it appears in Datadog.

To avoid instrumentation of function add comment directive anywhere in the file.

//instrument:exclude SomeFunc|SomeOtherfunc|privateFunc
...

func (s Cat) Name(ctx context.Context) (name string, err error) {
  //instrument:exclude Name

To instrument only specific functions add comment directive anywhere in the file and pass -all=false in CLI.

//instrument:include SomeFunc|SomeOtherfunc|privateFunc
...

func (s Cat) Name(ctx context.Context) (name string, err error) {
  //instrument:include Name

Functions that have named return err error will get spans with appropriate status and error recorded.

func (s Cat) Walk(ctx context.Context) (err error) {
  ...

Comments are supported through patching source files bytes and fmt.

Standard Go compiler directives are recognized. More details go help buildconstraint and https://pkg.go.dev/cmd/go#hdr-Build_constraints.

• //go:build exclude
• // +build exclude
• //go:build ignore
• // +build ignore

• Keeps comments
• Dynamic error variable name
• Dynamic ctx variable name
• Creating error when return is not named
• Detection if function is already instrumented
• Span Tags arguments
• Span Tags returns
• Changing _ to ctx when it is unused
• Mode to remove added instrumentation

It is laborious to add tracing code to every function manually. The code repeats 99% of time. Other languages can either modify code or have wrapper notations that makes even manual tracing much less laborious.

As of 2022-11-06, official Go does not support automatic function traces. https://go.dev/doc/diagnostics

Is there a way to automatically intercept each function call and create traces?

Go doesn’t provide a way to automatically intercept every function call and create trace spans. You need to manually instrument your code to create, end, and annotate spans.

Thus, providing automated version to add Trace Spans annotation.

Since we are adding multiple functions calls, it affects Go compiler decisions on inlining. It is expected that Go will less likely inline.

For example, can inline function

$ go build -gcflags="-m -m" ./internal/testdata 2>&1 | grep OneLine
internal/testdata/basic.go:80:6: can inline OneLineTypical with cost 62 as: func(context.Context, int) (int, error) { return fib(n), nil }

go-instrument -w -filename internal/testdata/basic.go

Can not inline after instrumentation

$ go build -gcflags="-m -m" ./internal/testdata 2>&1 | grep OneLine
internal/testdata/basic.go:132:6: cannot inline OneLineTypical: unhandled op DEFER

• https://github.com/hedhyw/otelinji — Very similar to current project. This tool gracefully handles code comments, so that its output can be tracked with normal code in version control. Main difference current project focuses on minimal code and dependencies.
• https://github.com/open-telemetry/opentelemetry-go-instrumentation — (in development) official eBPF based Go auto instrumentation
• https://github.com/keyval-dev/opentelemetry-go-instrumentation — eBPF based Go auto instrumentation of pre-selected libraries
• https://developers.mattermost.com/blog/instrumenting-go-code-via-ast — Very similar. Instrumenting Go code for tracing.
• https://github.com/gobwas/gtrace — non-OTEL, custom tracing framework that uses code generation

Java runtime modifies bytecode of methods on load time that adds instrumentation calls. Pre-defined libraries are instrumented (http, mysql, etc).

✅ Very short single line decorator statement can be used to trace selected methods.

Datadog

import datadog.trace.api.Trace

public class BackupLedger {
  @Trace
  public void write(List<Transaction> transactions) {
    for (Transaction transaction : transactions) {
      ledger.put(transaction.getId(), transaction);
    }
  }
}

OpenTelemetry

import io.opentelemetry.instrumentation.annotations.WithSpan;

public class MyClass {
  @WithSpan
  public void myMethod() {
      <...>
  }
}

✅ Automatic instrumentation of all functions is also possible.

Datadog supports wildcard for list of methods to trace.

dd.trace.methods
Environment Variable: DD_TRACE_METHODS
Default: null
Example: package.ClassName[method1,method2,...];AnonymousClass$1[call];package.ClassName[]
List of class/interface and methods to trace. Similar to adding @Trace, but without changing code. Note: The wildcard method support ([]) does not accommodate constructors, getters, setters, synthetic, toString, equals, hashcode, or finalizer method calls

java -javaagent:/path/to/dd-java-agent.jar -Ddd.service=web-app -Ddd.env=dev -Ddd.trace.methods="*" -jar path/to/application.jar

• Java Auto-Instrumentation
• Datadog Java Auto-Instrumentation
• Datadog Java Tracing Config
• Datadog Instrumentation Business Logic
• Javaassist

Python monkeypatching of functions at runtime is used to add instrumentation calls. Pre-defined libraries are instrumented (http, mysql, etc).

✅ Very short single line decorator statement can be used to trace selected methods.

Datadog

from ddtrace import tracer

class BackupLedger:
    @tracer.wrap()
    def write(self, transactions):
        for transaction in transactions:
            self.ledger[transaction.id] = transaction

OpenTelemetry

@tracer.start_as_current_span("do_work")
def do_work():
    print("doing some work...")

⚠️ Automatic instrumentation of all functions is also possible via monkeypatching (fidning stable library is pending).

• OpenTelemetry Python Instrumentation
• Blog: Timescale: OpenTelemetry and Python: A Complete Instrumentation Guide
• https://github.com/harshitandro/Python-Instrumentation

❌ Only manual instrumentation.

✅ Very short single line decorator statement can be used to trace selected functions with well-establisehd tokio framework.

#[tracing::instrument]
pub fn shave(yak: usize) -> Result<(), Box<dyn Error + 'static>> {

#[instrument]
async fn write(stream: &mut TcpStream) -> io::Result<usize> {

• https://github.com/tokio-rs/tracing
• https://opentelemetry.io/docs/instrumentation/rust
• https://docs.rs/opentelemetry/latest/opentelemetry
• https://github.com/open-telemetry/opentelemetry-rust/tree/main/examples
• https://docs.rs/datadog-apm/latest/datadog_apm

With eBPF we can track latency, but we would not be able to assign errors to spans. Some platforms may not have access to eBPF.

Benefit of wrapping is to keep original code without modifications. However, manual step for switching would still be requied. Given every single function is duplciated and is within same package, code will quickly become messy and hard to maintain by user.

Typically, packages are failry big and performs lots of logic. Oftencase, business domains are split only in few large packages. Low level packages are already likely to be traced with standard tracing (MySQL, het/http, etc). Thus, it is doubtful how much benefit would be from tracing only exported functions and only on import.

This would lead to circular dependency failure, since some even exported functions in original package may be called withing same package. Thus, we would either skip those calls, or fail with circular dependency while trying to wrap those.

1.97K spans, fibbonaci

3.7K spans, go cover treemap

Orignal version was using go:instrument directive. However, many members of Go community raised concern that it takes over reserved core Go toolchain directives (eg, //go:norace). Even though as of 2022-11-25 Go core does not use go:instrument, to respect community and Go core, leaving using //instrument: directive instead.

One of proposed solutions for selectors was to use regex.

Specifically, first usecase proposed was to use

//instrument:exclude .*
//instrument:include ^API.*$

The issue with this is collision of two functions:

• A) exlude all and select specific
• B) include all and exclude specific

Similarly, there is collision of subspace of functions for exclusion and inclusion.

As of 2022-11-25, @nikolaydubina does not know how to resolve this better. Thus, keeping simple map matching wiht and condition of overlaps.