Metrics are a powerful and cost-efficient tool for understanding the health and performance of your code in production, but it's hard to decide what metrics to track and even harder to write queries to understand the data.
Autometrics is a Go Generator bundled with a library that instruments your functions with the most useful metrics: request rate, error rate, and latency. It standardizes these metrics and then generates powerful Prometheus queries based on your function details to help you quickly identify and debug issues in production.
- โจ
//autometrics:inst
directive adds useful metrics to any function, without you thinking about what metrics to collect - ๐ก Generates powerful Prometheus queries to help quickly identify and debug issues in production
- ๐ Injects links to live Prometheus charts directly into each function's doc comments
- ๐ Grafana dashboards work without configuration to visualize the performance of functions & SLOs
- ๐ Correlates your code's version with metrics to help identify commits that introduced errors or latency
- ๐ Standardizes metrics across services and teams to improve debugging
- โ๏ธ Function-level metrics provide useful granularity without exploding cardinality
- ๐จ Define alerts using SLO best practices directly in your source code
- ๐ Attach exemplars automatically to connect metrics with traces
- โ๏ธ Configurable metric collection library
(
opentelemetry
orprometheus
)
See autometrics.dev for more details on the ideas behind autometrics.
When alerting rules are added, code annotations make Prometheus trigger alerts directly from production usage:
A fully working use-case and example of library usage is available in the examples/web subdirectory. You can build and run load on the example server using:
git submodule update --init
docker compose -f docker-compose.prometheus-example.yaml up
And then explore the generated links by opening the main file in your editor.
There is a one-time setup phase to prime the code for autometrics. Once this
phase is accomplished, only calling go generate
is necessary.
The generator is the binary in cmd/autometrics, so the easiest way to get it is to install it through go:
go install github.com/autometrics-dev/autometrics-go/cmd/autometrics@latest
Make sure your `$PATH` is set up
In order to have `autometrics` visible then, make sure that the directory `$GOBIN` (or the default `$GOPATH/bin`) is in your `$PATH`:$ echo "$PATH" | grep -q "${GOBIN:-$GOPATH/bin}" && echo "GOBIN in PATH" || echo "GOBIN not in PATH, please add it"
GOBIN in PATH
In the main entrypoint of your program, you need to both add package
import (
"github.com/autometrics-dev/autometrics-go/prometheus/autometrics"
)
And then in your main function initialize the metrics
shutdown, err := autometrics.Init()
if err != nil {
log.Fatalf("could not initialize autometrics: %s", err)
}
defer shutdown(nil)
Init
takes optional arguments to customize the metrics. The main ones are WithBranch
,
WithService
, WithVersion
, and WithCommit
; it will add relevant information on the
metrics for better intelligence:
shutdown, err := autometrics.Init(
autometrics.WithService("myApp"),
autometrics.WithVersion("0.4.0"),
)
if err != nil {
log.Fatalf("could not initialize autometrics: %s", err)
}
defer shutdown(nil)
You can use any string variable whose value is
injected at build time by ldflags
for example, or use environment variables.
Note Instead of hardcoding the service in the code, you can simply have environment variables set to fill the "Service" name.
AUTOMETRICS_SERVICE_NAME
will be used if set, otherwiseOTEL_SERVICE_NAME
will be attempted (so OpenTelemetry compatibility comes out of the box).
If you have am
installed in version 0.6.0
or later, you can
use am instrument single -e /vendor/ -l go .
to instrument everything (excluding a possible /vendor
subdirectory)
Use find and sed to insert a //go:generate
directive that will instrument all
the functions in all source files under the current directory:
(Replace gsed
with sed
on linux; gsed
is installed with brew gsed
)
find . \
-type d -name vendor -prune -or \
-type f -name '*.go' \
-print0 | xargs -0 gsed -i -e '/package/{a\//go:generate autometrics --inst-all --no-doc' -e ':a;n;ba}'
You can remove the --no-doc
to get the full experience, but the generator will add a lot of comments if so.
This grants you more control over what gets instrumented, but it is longer to add.
Warning You must both add the
//go:generate
directive, and one//autometrics:inst
directive per function you want to instrument
On top of each file you want to use Autometrics in, you need to have a go generate
cookie:
//go:generate autometrics
Then instrumenting functions depend on their signature, expand the corresponding subsection to see details:
- if the function returns an
error
, or - if the function is a
http.Handler
.
Once it is done, you can call the generator
Expand to instrument error returning functions
Given a starting function like:
func AddUser(args any) error {
// Do stuff
return nil
}
The manual changes you need to do are:
+//autometrics:inst
-func AddUser(args any) error {
+func AddUser(args any) (err error) {
// Do stuff
return nil
}
The generated metrics will count a function as having failed if the err
return value is non-nil.
Warning If you want the generated metrics to contain the function success rate, you must name the error return value. This is why we recommend to name the error value you return for the function you want to instrument.
Expand to instrument HTTP handlers functions
Autometrics comes with a middleware library for net.http
handler functions.
- Import the middleware library
import "github.com/autometrics-dev/autometrics-go/prometheus/midhttp"
- Wrap your handlers in
Autometrics
handler
http.Handle(
"/path",
+ midhttp.Autometrics(
- http.HandlerFunc(routeHandler),
+ http.HandlerFunc(routeHandler),
+ // Optional: override what is considered a success (default is 100-399)
+ autometrics.WithValidHttpCodes([]autometrics.ValidHttpRange{{Min: 200, Max: 299}}),
+ // Optional: Alerting rules
+ autometrics.WithSloName("API"),
+ autometrics.WithAlertSuccess(90),
+ )
)
The generated metrics here will count a function as having failed if the return
code of the handler is bad (in the 4xx
and 5xx
ranges). The code snippet
above shows how to override the ranges of codes that should be considered as
errors for the metrics/monitoring.
Note There is only middleware for
net/http
handlers for now, but support for other web frameworks will come as needed/requested! Don't hesitate to create issues in the repository.
Warning To properly report the function name in the metrics, the autometrics wrapper should be the innermost middleware in the stack.
You can now call go generate
:
$ go generate ./...
The generator will augment your doc comment to add quick links to metrics (using the Prometheus URL as base URL), and add a unique defer statement that will take care of instrumenting your code.
autometrics --help
will show you all the different arguments that can control
behaviour through environment variables. The most important options are
changing the
target of
generated links, or disabling doc generation to
keep only instrumentation
The last step now is to actually expose the generated metrics to the Prometheus instance.
The shortest way is to reuse prometheus/promhttp
handler in your main entrypoint:
import (
"github.com/autometrics-dev/autometrics-go/prometheus/autometrics"
"github.com/prometheus/client_golang/prometheus/promhttp"
)
func main() {
shutdown, err := autometrics.Init(
autometrics.WithVersion("0.4.0"),
autometrics.WithCommit("anySHA"),
autometrics.WithService("myApp"),
)
http.Handle("/metrics", promhttp.Handler())
}
This is the shortest way to initialize and expose the metrics that autometrics will use in the generated code.
A Prometheus server can be configured to poll the application, and the autometrics will be available! (See the Web App example for a simple, complete setup)
You can use the open source Autometrics CLI to run automatically configured Prometheus locally to see the metrics that will be registered by the change. See the Autometrics CLI docs for more information.
or you can configure Prometheus manually:
scrape_configs:
- job_name: my-app
metrics_path: /metrics # the endpoint you configured your metrics exporter on (usually /metrics)
static_configs:
- targets: ['localhost:<PORT>'] # The port your service is on
scrape_interval: 200ms
# For a real deployment, you would want the scrape interval to be
# longer but for testing, you want the data to show up quickly
You can also check the documentation to find out about setting up Prometheus locally, with Fly.io, or with Kubernetes
Change the annotation of the function to automatically generate alerts for it:
//autometrics:inst --slo "Api" --success-target 90
func AddUser(args any) (err error) {
// Do stuff
return nil
}
Then you need to add the bundled recording rules to your prometheus configuration.
The valid arguments for alert generation are:
--slo
(MANDATORY for alert generation): name of the service for which the objective is relevant--success-rate
: target success rate of the function, between 0 and 100 (you must name theerror
return value of the function for detection to work.)--latency-ms
: maximum latency allowed for the function, in milliseconds.--latency-target
: latency target for the threshold, between 0 and 100 (so X% of calls must last less thanlatency-ms
milliseconds). You must specify both latency options, or none.
Warning The generator will error out if you use percentile targets that are not supported by the bundled Alerting rules file. Support for custom target is planned but not present at the moment
Warning You MUST have the
--latency-ms
values to match the values given in the buckets given in theautometrics.Init
call. The values in the buckets are given in seconds. By default, the generator will error and tell you the valid default values if they don't match. If the default values inautometrics.DefBuckets
do not match your use case, you can change the buckets in the init call, and add a--custom-latency
argument to the//go:generate
invocation.
-//go:generate autometrics
+//go:generate autometrics --custom-latency
When using the Prometheus library for metrics collection, it automatically adds trace and span information in the metrics as exemplars that can be queried with Prometheus, if the server is configured correctly
Autometrics supports using OpenTelemetry with a prometheus exporter instead of using Prometheus to publish the metrics. The changes you need to make are:
- change where the
autometrics
import points to
import (
- "github.com/autometrics-dev/autometrics-go/prometheus/autometrics"
+ "github.com/autometrics-dev/autometrics-go/otel/autometrics"
)
- maybe change the call to
autometrics.Init
to the new signature: instead of a registry, theInit
function takes a meter name for theotel_scope
label of the exported metric. That meansautometrics
won't have aWithRegistry
option anymore, but aWithMeterName
instead.
shutdown, err := autometrics.Init(
- autometrics.WithRegistry(nil),
+ autometrics.WithMeterName("myApp/v2/prod"),
autometrics.WithVersion("2.1.37"),
autimetrics.WithCommit("anySHA"),
autometrics.WithService("myApp"),
)
- add the
--otel
flag to the//go:generate
directive
-//go:generate autometrics
+//go:generate autometrics --otel
Why would I use a push-based workflow?
If you have an auto-scaled service (with instances spinning up and down), maintaining the configuration/discovery of instances on the Prometheus side of things can be hard. Using a push-based workflow inverts the burden of configuration: all your instances generate a specific ID, and they just need to push metrics to a given URL. So the main advantages of a push-based workflow appear when the the set of machines producing metrics is dynamic:
- Your Prometheus configuration does not need to be dynamic anymore, it's "set and forget" again
- No need to configure service discovery separately (which can be error-prone)
It can be summarized with one sentence. The monitoring stack (Prometheus/OpenTelemetry collector) does not need to know the infrastructure of application deployment; nor does the application code need to know the infrastructure of the monitoring stack. Decoupling prevents configuration-rot.
If you don't want to/cannot configure your Prometheus instance to scrape the instrumented code, Autometrics provides a way to push metrics instead of relying on a polling collection process.
Note It is strongly advised to use the OpenTelemetry variant of Autometrics to support push-based metric collection. Prometheus push gateways make aggregation of data across multiple sources harder.
How can I use a push-based workflow with Autometrics?
If you have a Prometheus push
gateway or an OTLP
collector setup with an accessible
URL, then you can directly switch from metric polling to metric pushing by
passing the push-related options to autometrics.Init
:
shutdown, err := autometrics.Init(
autometrics.WithMeterName("myApp/v2/prod"),
autometrics.WithVersion("2.1.37"),
autometrics.WithService("myApp"),
+ autometrics.WithPushCollectorURL("https://collector.example.com"),
+ autometrics.WithPushJobName("instance_2"), // You can leave the JobName out to let autometrics generate one
+ autometrics.WithPushPeriod(1 * time.Second), // Period is only relevant (and available) when using OpenTelemetry implementation
+ autometrics.WithPushTimeout(500 * time.Millisecond), // Timeout is only relevant (and available) when using OpenTelementry implementation
)
Note If you do not want to setup an OTLP collector or a Prometheus push-gateway yourself, you can contact us so we can setup a managed instance of Prometheus for you. We will effectively give you collector URLs, that will work with both OpenTelemetry and Prometheus; and can be visualized easily with our explorer as well!
Monitoring/Observability must not crash the application.
So when Autometrics encounters
errors, instead of bubbling it up until the program stops, it will log the error and absorb it.
To leave choice in the logging implementation (depending on your application dependencies),
Autometrics exposes a Logger
interface you can implement and then inject in the Init
call
to have the logging you want.
The Logger
interface is a subset of slog.Logger
methods, so that most loggers can be used.
Autometrics also provides 2 simple loggers out of the box:
NoOpLogger
, which is the default logger and does nothing,PrintLogger
which usesfmt.Print
to write logging messages on stdout.
To use the PrintLogger
instead of the NoOpLogger
for examble, you just have to change
the Init
call:
shutdown, err := autometrics.Init(
autometrics.WithMeterName("myApp/v2/prod"),
autometrics.WithVersion("2.1.37"),
autometrics.WithService("myApp"),
+ autometrics.WithLogger(autometrics.PrintLogger{}),
)
As autometrics is a Go generator that modifies the source code when run, it
might be interesting to set up go generate ./...
to run in a git pre-commit
hook so that you never forget to run it if you change the source code.
If you use a tool like pre-commit, see their
documentation about how to add a hook that will run go generate ./...
.
Otherwise, a simple example has been added in the configs folder
as an example. You can copy this file in your copy of your project's repository, within
.git/hooks
and make sure that the file is executable.
By default, the generated links will point to localhost:9090
, which the default location
of Prometheus when run locally.
The environment variable AM_PROMETHEUS_URL
controls the base URL of the instance that
is scraping the deployed version of your code. Having an environment variable means you
can change the generated links without touching your code. The default value, if absent,
is http://localhost:9090/
.
You can have any value here, the only adverse impact it can have is that the links in the doc comment might lead nowhere useful.
By default, autometrics will add a lot of documentation on each instrumented function. If you prefer not having the extra comments, but keep the instrumentation only, you have multiple options:
- To disable documentation on a single function, add the
--no-doc
argument to the//autometrics:inst
directive:
-//autometrics:inst
+//autometrics:inst --no-doc
- To disable documentation on a file, add the
--no-doc
argument to the//go:generate
directive:
-//go:generate autometrics
+//go:generate autometrics --no-doc
- To disable documentation globally, use the environment variable
AM_NO_DOCGEN
:
$ AM_NO_DOCGEN=true go generate ./...
If for some reason you want to stop using autometrics, the easiest way includes 3 steps.
First is to use the environment variable AM_RM_ALL
to remove all generated documentation
and code:
$ AM_RM_ALL=true go generate ./...
Second step is to use grep/your text-editor/your IDE of choice to remove all lines starting with
//go:generate autometrics
from your files so go generate
will stop creating autometrics code.
Last step is to use your text-editor IDE to remove remnants:
- a "go imports" cleaner to remove the
autometrics
imports (or grep-ing again) - remove the
autometrics.Init
call from the main entrypoint of your code.
The first version of the library has not been written by Go experts. Any comment or code suggestion as Pull Request is more than welcome!
Issues, feature suggestions, and pull requests are very welcome!
If you are interested in getting involved:
- Join the conversation on Discord
- Ask questions and share ideas in the Github Discussions
- Take a look at the overall Autometrics Project Roadmap