lambda-framework-go

README

lambda-framework-go

A small framework for AWS Lambda functions written in Go, on Powertools for AWS Lambda (Go): what AWS recommends, done the way AWS recommends it, in a comfortable shape. main.go configures; each handler receives its clients and does its job.

It wires the Powertools logger, tracer and metrics, the Powertools Event Handler router with typed, validated operations, the AWS SDK clients and a database pool, to Postgres or to Aurora DSQL, around your business handlers, and gives every failure one place to be reported.

Requires Go 1.27.

go get github.com/alessandrolattao/lambda-framework-go

Packages

Package	What it does
framework	Container with the dependencies; Get/Post/Put/Patch/Delete and HTTP for API Gateway HTTP APIs; Handle and Consume for every other event; error and panic reporting
request	What typed binding does not cover: the JWT claims API Gateway verified, the client IP behind trusted proxies, comma-separated id lists
response	The error envelope, success responses with cache headers, CSV and content negotiation
migrate	The business handler of a migration Lambda: goose SQL migrations on Aurora DSQL (or Postgres)

Quick start

An HTTP API

package main

import (
	"context"
	"errors"
	"uuid"

	"github.com/aws/aws-lambda-go/lambda"
	"github.com/jackc/pgx/v5"

	"github.com/alessandrolattao/lambda-framework-go/framework"
	"github.com/alessandrolattao/lambda-framework-go/response"
)

func main() {
	c := framework.New()
	c.Get("/orders/{id}", getOrder)
	lambda.Start(c.HTTP())
}

// orderID is what the route reads; the router binds and validates it.
type orderID struct {
	ID uuid.UUID `path:"id" format:"uuid"`
}

type order struct {
	ID    uuid.UUID `json:"id"`
	Total int64     `json:"total"`
}

func getOrder(ctx context.Context, c *framework.Container, in orderID) (order, error) {
	o := order{ID: in.ID}
	err := c.DB.QueryRow(ctx, "SELECT total FROM orders WHERE id = $1", in.ID).Scan(&o.Total)
	if errors.Is(err, pgx.ErrNoRows) {
		return order{}, response.NotFound("orderNotFound", "no order with this id", nil)
	}
	if err != nil {
		return order{}, response.InternalServerError(err, "serverError", "could not read the order", nil)
	}
	return o, nil
}

Input structs follow the Powertools Event Handler: fields tagged path, query, header, cookie, a Body field decoded from JSON, constraints as tags (minimum, maxLength, pattern, enum, format, required...). A request that does not conform never reaches the handler: it gets a 422. The same types describe the API: c.Router().OpenAPIJSON() is its OpenAPI document.

Any other event

Handlers that only succeed or fail (EventBridge, SQS, S3, CloudWatch Logs) use Consume; handlers that answer with a result use Handle:

lambda.Start(c.Consume(func(ctx context.Context, c *framework.Container, e events.SQSEvent) error {
	// ...
	return nil
}))

The other Powertools utilities (batch, idempotency, parameters, feature flags, validation, parser...) work inside these handlers as they do anywhere.

What happens around a handler

• The Powertools decorators, stacked as Powertools recommends. The logger's InjectContext (Lambda context keys, correlation id for API Gateway, ALB, Function URL and EventBridge events, sampling, buffering), the tracer's CaptureLambdaHandler (the invocation span, ColdStart and Service annotations) and, when metrics are configured, LogMetrics (one metric set per invocation, flushed at its end).
• Initialization once per cold start. The AWS config, the clients (S3, EventBridge, SFN, Dynamo, Websocket) and the database pool (Postgres or Aurora DSQL) are created on the first invocation and reused afterwards. A failed initialization is retried by the next invocation. A client you put in a field before that (a fake, in a test) is kept.
• Per-invocation state in the context. The handler shares the container, which is read-only once initialized; the request id, the client's address (client_ip) and the metric set live in the context. Log with the Context methods (c.Logger.InfoContext(ctx, ...)) and the records carry them, also when one execution environment runs several invocations at once (Lambda Managed Instances).
• One door for failures. Return the error and log nothing: the framework writes it once, at Error, with the request id. Wrap the context only the handler knows into the error (fmt.Errorf("charge order %v: %w", id, err)). A panic is written down with its stack.
• Failures Lambda can tell apart. A failure reaches Lambda (and Step Functions, which matches Retry and Catch rules on it) with an errorType that means something: the type of the application's or library's error in the chain (PgError, ...), or HandlerError, InvalidEvent, InvalidResult, InitError. aws-lambda-go's own plain text copy of the failure goes to Debug.
• Every API request leaves a line. API request completed carries the status, the route and the duration, so a 404 or a 422 is visible too.
• API Gateway always gets an answer, in one envelope. A *response.Error answers with {"code","message","details","timestamp"}; so do the router's own errors: an unknown route (routeNotFound), a request that fails validation (invalidParam, invalidBody, with the Powertools details), an unsupported content type, a Powertools ServiceError. A failure no handler answered becomes a 500 serverError. A 5xx is reported once, with its cause, which never reaches the client. Lambda counts such invocations as successes, so alarm on the error logs rather than on the Errors metric.
• JSON v2 both ways. Events are decoded and results encoded with encoding/json/v2; response bodies with sorted map keys.
• Tracing with OpenTelemetry, through the Powertools tracer: the X-Ray UDP exporter the AWS X-Ray migration guide prescribes for Lambda. Each invocation is a server span under the trace Lambda started, the handler runs in a ## businessLogic span, and the AWS SDK calls and the database queries (through otelpgx: operation and SQL text, never the parameters) made with its context are spans below it. An API request is described on the invocation's span (X-Ray's http block, http_route and http_status annotations); a 5xx marks both spans as failed. Open your own spans with c.Tracer.Capture(ctx, name, fn). Turn on active tracing on the function: without an X-Ray endpoint, tracing is off.

Configuration

Options of framework.New:

Option	Default	Meaning
WithService(name)	POWERTOOLS_SERVICE_NAME, else the function name	The service of the logs, the traces and the metrics
WithLogging(opts...)		Options of the Powertools logger (level, sampling, buffering, format)
WithTracing(opts...)		Options of the Powertools tracer
WithTracerProvider(tp)	X-Ray UDP exporter	Another OpenTelemetry tracer provider
WithMetrics(opts...)	off	Turns the Powertools metrics on (namespace, dimensions, Datadog)
WithRouter(opts...)		Options of the Event Handler router (CORS, debug, strip prefixes, OpenAPI)
WithDatabaseURL(url)	DATABASE_URL	Postgres connection string
WithDSQLEndpoint(endpoint)	DSQL_ENDPOINT	Aurora DSQL cluster endpoint, in the region in AWS_REGION
WithTrustedProxies(n)	0	Proxies in front of API Gateway (CloudFront is 1), used to find the client IP

Environment variables: every POWERTOOLS_* variable of the utilities (POWERTOOLS_SERVICE_NAME, POWERTOOLS_LOG_LEVEL, POWERTOOLS_LOGGER_SAMPLE_RATE, POWERTOOLS_METRICS_NAMESPACE, POWERTOOLS_TRACE_DISABLED, POWERTOOLS_DEV for indented logs while developing...), Lambda's AWS_LAMBDA_LOG_LEVEL, and:

Variable	Meaning
DATABASE_URL	Postgres connection string, when WithDatabaseURL is not used
DSQL_ENDPOINT	Aurora DSQL cluster endpoint, when WithDSQLEndpoint is not used
AWS_REGION	The region of the Aurora DSQL cluster; Lambda sets it to the function's own
WS_ENDPOINT	WebSocket API endpoint; without it Websocket is nil
AWS_ENDPOINT_URL, AWS_ENDPOINT_URL_S3	A custom S3 endpoint (MinIO, LocalStack, Floci); S3 then uses path-style addressing

Metrics are on when WithMetrics is used or POWERTOOLS_METRICS_NAMESPACE is set; without them no invocation flushes an empty metric set.

A container has at most one database: Postgres with a connection string, Aurora DSQL with a cluster endpoint. With neither there is no pool and DB is nil; with both, the first invocation fails. Either way DB is a *pgxpool.Pool, and handlers use it the same way.

Postgres

The Postgres pool holds one connection per invocation the execution environment can run at once, unless the connection string sets pool_max_conns. Every statement is capped to the time the invocation has left.

Aurora DSQL

DSQL has no password: a connection authenticates with an IAM token. The pool goes through AWS's own connector, Aurora DSQL Connector for Go on pgx, which signs a fresh token, valid 15 minutes, for every new connection with the function's credentials. A warm execution environment that opens a connection long after its cold start still gets in.

• The pool connects to the postgres database as the database user admin, over TLS verified against the endpoint. The function's role needs dsql:DbConnectAdmin on the cluster.
• It holds one connection per invocation the execution environment can run at once, and recycles them with the connector's own lifetimes (55 minutes, 10 minutes idle), below the hour after which DSQL closes a connection.
• Statements are not capped server side: DSQL rejects statement_timeout and ends every transaction after 5 minutes on its own. A query still running when the invocation's deadline passes fails at that deadline, on the client.
• Queries run in pgx's QueryExecModeDescribeExec, as the connector sets it: safe across live schema changes, at the cost of one more round trip per query.
• DSQL resolves conflicting transactions with optimistic concurrency control: the transaction that commits second fails with SQLSTATE OC000, OC001 or 40001, and is meant to be retried. The connector's occretry package retries a transaction or a statement on c.DB.

With SST

SST encrypts linked resources into the function bundle, so the database password never sits in a plaintext environment variable. Keep it that way: read the linked resource in main and pass the connection string.

Postgres

import (
	"fmt"
	"net"
	"net/url"

	"github.com/sst/sst/v3/sdk/golang/resource"
)

func databaseURL() (string, error) {
	get := func(key string) (string, error) {
		v, err := resource.Get("Database", key)
		if err != nil {
			return "", err
		}
		return fmt.Sprint(v), nil
	}
	host, err := get("host")
	if err != nil {
		return "", err
	}
	port, _ := get("port")
	user, _ := get("username")
	password, _ := get("password")
	name, _ := get("database")

	u := url.URL{
		Scheme: "postgres",
		User:   url.UserPassword(user, password),
		Host:   net.JoinHostPort(host, port),
		Path:   "/" + name,
	}
	return u.String(), nil
}

func main() {
	dsn, err := databaseURL()
	if err != nil {
		panic(err)
	}
	c := framework.New(framework.WithDatabaseURL(dsn))
	c.Get("/orders/{id}", getOrder)
	lambda.Start(c.HTTP())
}

Aurora DSQL

Link the cluster to the function: the link carries the endpoint and grants the function's role dsql:DbConnect and dsql:DbConnectAdmin on the cluster.

const database = new sst.aws.Dsql("Database");

new sst.aws.Function("Orders", {
  handler: "./cmd/orders",
  runtime: "go",
  link: [database],
});

Read the endpoint in main and pass it on; the region is the function's own, in AWS_REGION.

import (
	"fmt"

	"github.com/sst/sst/v3/sdk/golang/resource"
)

func main() {
	endpoint, err := resource.Get("Database", "endpoint")
	if err != nil {
		panic(err)
	}
	c := framework.New(framework.WithDSQLEndpoint(fmt.Sprint(endpoint)))
	c.Get("/orders/{id}", getOrder)
	lambda.Start(c.HTTP())
}

Migrations

Each database gets a migration Lambda: a main package that embeds its goose SQL files and hands migrate.Up to Handle. The code that runs them is in the framework; the Lambda brings only its SQL.

package main

import (
	"embed"

	"github.com/aws/aws-lambda-go/lambda"

	"github.com/alessandrolattao/lambda-framework-go/framework"
	"github.com/alessandrolattao/lambda-framework-go/migrate"
)

//go:embed *.sql
var migrations embed.FS

func main() {
	c := framework.New()
	lambda.Start(c.Handle(migrate.Up(migrations)))
}

It is a business handler like any other: it gets the container's logging, tracing, error reporting and pool, and connects as the container does, to DSQL with WithDSQLEndpoint or DSQL_ENDPOINT, in the region in AWS_REGION, as admin with an IAM token. Every invocation, whatever its payload, applies the migrations the database has not applied yet, up to the latest, and answers with what it did:

{"version": 3, "applied": [2, 3]}

An invocation with nothing new applies nothing and succeeds. goose logs through the container's logger, at Info, every statement it runs and every migration it applies (logger=goose): the logs of a failed deploy show its last statement. Migrations go up only: no Down section runs, and there are no Go migrations. migrate is a package of its own, which framework does not import: goose is linked into the migration Lambdas only, and the other Lambdas keep their size and their cold start.

What DSQL asks of a migration:

• -- +goose NO TRANSACTION in every file. A DSQL transaction holds at most one DDL statement, and never DDL together with DML. Without a transaction around the file, each statement is a transaction of its own. Up checks every file before it applies anything, and fails naming the files without the annotation.

  -- +goose NO TRANSACTION
  -- +goose Up
  CREATE TABLE orders (id uuid PRIMARY KEY, total bigint NOT NULL);
  CREATE INDEX ASYNC orders_total ON orders (total);
  

• CREATE INDEX ASYNC is how DSQL creates an index: it returns at once and builds the index in a job. After each migration Up waits, with sys.wait_for_job, for every asynchronous job not yet completed, so the next migration finds the schema complete; a job that ends failed fails the invocation, naming the job and the migration.

• Conflicts are retried. A statement that loses an optimistic concurrency conflict (SQLSTATE 40001, OC000, OC001) committed nothing, and runs again, up to 5 times, after a random wait of at most 100 ms, then 200 ms, doubling, never more than 3 seconds. A migration that still fails fails the invocation: the migrations before it stay applied, and the next invocation starts from the one that failed. So does the deploy that invoked it.

• The role needs dsql:DbConnectAdmin on the cluster: migrations change the schema as admin.

• The timeout covers the job waits. An index build on a large table takes minutes: give the function a timeout above the longest one (Lambda allows up to 15 minutes).

• Runs never overlap. There is no lock (DSQL has no advisory locks): invoke the migration Lambda once per deploy and wait for its answer before the next deploy.

On Postgres the same Lambda works with DATABASE_URL; there is no sys.jobs there, and the job wait is skipped.

With SST

Link the cluster to the migration function, as to any other, and give it a timeout for the index builds. Invoking it on each deploy is up to your infrastructure.

const database = new sst.aws.Dsql("Database");

new sst.aws.Function("OrdersMigrations", {
  handler: "./cmd/orders-migrations",
  runtime: "go",
  link: [database],
  timeout: "15 minutes",
});

func main() {
	endpoint, err := resource.Get("Database", "endpoint")
	if err != nil {
		panic(err)
	}
	c := framework.New(framework.WithDSQLEndpoint(fmt.Sprint(endpoint)))
	lambda.Start(c.Handle(migrate.Up(migrations)))
}

Testing a handler

Build a Container by hand with fakes in its fields and call the business handler directly: a hand-built container is never initialized behind your back, and it does not trace.

c := &framework.Container{Logger: logging.New(logging.WithWriter(io.Discard)), EventBridge: fakeBus}
got, err := getOrder(t.Context(), c, orderID{ID: id})

To go through the framework as well, routing and validation included, call the Lambda handler with an event: c.HTTP().Call(ctx, request), or c.Handle(h).Call(ctx, event).

Development

go test ./...
golangci-lint run ./...

The lint configuration is strict, the same as Powertools'. The Aurora DSQL tests need neither a cluster nor the network: the pool connects lazily, and they sign tokens with fixed credentials. The Postgres tests skip themselves without a database. To run them:

docker run -d --rm -e POSTGRES_PASSWORD=postgres -p 55432:5432 postgres:18-alpine
TEST_DATABASE_URL='postgres://postgres:postgres@localhost:55432/postgres?sslmode=disable' go test ./...

The migration tests run goose on that Postgres, each in a schema of its own, and skip without TEST_DATABASE_URL. One of them needs a real Aurora DSQL cluster, because only DSQL has CREATE INDEX ASYNC and its jobs: it runs with the cluster endpoint in TEST_DSQL_ENDPOINT and AWS credentials allowed dsql:DbConnectAdmin on it, and skips without them.

The parsers that read what a client sends have fuzz tests; go test runs their seeds, and fuzzing one goes like this:

go test -run '^$' -fuzz '^FuzzClientIP$' -fuzztime 1m ./request

End-to-end tests

e2e/ is a module of its own (so its test dependencies stay out of this one) that runs two real Lambdas, an API Gateway HTTP API one and an EventBridge one, the way Lambda does: built for provided.al2023, run in the Lambda base image under the Runtime Interface Emulator, with the memory and the CPU share of a Lambda size. Postgres runs next to them (Aurora DSQL has no local version, so the suite runs on Postgres), and a UDP listener stands in for the X-Ray endpoint, so the tests check responses, database writes, log records and exported spans, on the happy paths and the failing ones. It needs Docker.

cd e2e
go test -run TestOrders ./...   # behaviour, about 15 s
go test ./...                   # plus startup times at 128, 512 and 1024 MB, about 100 s

TestStartup measures five cold starts per size and fails when the median cold invocation goes over its limit. E2E_MAX_COLD_START_SCALE=2 doubles the limits on a slower machine; E2E_STARTUP_REPORT=startup.md writes the table of measurements to a file.

License

MIT, see LICENSE.