Esaldgut AI Native Engineering Workspace Lambda Go Lazy Init Segregated

Segregated lazy init for Go Lambdas

AWS Lambda's best practices say verbatim: "Initialize SDK clients and database connections outside of the function handler … Subsequent invocations processed by the same instance of your function can reuse these resources." True — but the language-neutral doc stops short of the Go-specific refinement: in Go, anything you initialize during INIT (an init() block or an eager package-level global) runs unconditionally on every cold start, on every code path. init() cascades across all imported packages. So a Mongo connection opened in init() taxes 100% of cold starts even if 70% of invocations never touch Mongo.

The fix: defer each expensive client behind its own sync.Once, grouped by temporal locality of use — not one initAll().

When to invoke

• A Go Lambda has 2+ clients (AWS SDK service clients, Mongo, an HTTP client to a slow upstream) used on different branches.
• You see init() (or an eager global) opening a database/Mongo connection or constructing a client the common path doesn't use.
• Adding a new client to an existing handler — wire it to the branch that needs it, not to init().

Announce on invoke: "Using lambda-go-lazy-init-segregated to split sync.Once by usage path so the happy path doesn't pay cold-start for unused clients."

The mental model

• The AWS aws.Config itself is cheap — load it once (globally or behind one sync.Once).
• Each serviceX.NewFromConfig(cfg) is cheap-ish; a Mongo dial or a TLS handshake to a slow upstream is the expensive part.
• Group sync.Once by which branch needs the client together, not "all infra clients in one Once."
• sync.Once.Do guarantees exactly-once, goroutine-safe, lazy execution — the idiomatic deferral primitive.

Example shape: an inbound-federation handler with 3 clients (an identity provider, EventBridge, Mongo) across 2 segregated Once values. The most common branch — a brand-new federated sign-up with no candidate to reconcile — initializes only the identity-provider client, never Mongo or EventBridge.

Why Go specifically

Java/Python/.NET have SnapStart, which snapshots a pre-initialized environment so eager init is effectively free. Go has no SnapStart. Every cold start re-runs INIT from scratch, so segregated lazy init is the primary lever for Go cold-start cost. (If you do use Provisioned Concurrency, INIT runs during pre-warm and the value diminishes — this pattern matters most for on-demand Lambdas with bursty traffic and seldom-used branches.)

Canonical example

package main

import (
	"context"
	"sync"

	"github.com/aws/aws-lambda-go/lambda"
	"github.com/aws/aws-sdk-go-v2/aws"
	"github.com/aws/aws-sdk-go-v2/config"
	"github.com/aws/aws-sdk-go-v2/service/eventbridge"
	"github.com/aws/aws-sdk-go-v2/service/cognitoidentityprovider"
	"go.mongodb.org/mongo-driver/v2/mongo"
)

// aws.Config: loaded once, cheap. Shared by all service clients.
var (
	cfgOnce sync.Once
	awsCfg  aws.Config
)

func loadAWSConfig(ctx context.Context) aws.Config {
	cfgOnce.Do(func() {
		c, err := config.LoadDefaultConfig(ctx)
		if err != nil {
			panic(err) // init failure: fail fast, structured error to Lambda
		}
		awsCfg = c
	})
	return awsCfg
}

// Once #1 — identity path. Every invocation needs the IdP client (cheap).
var (
	idpOnce   sync.Once
	idpClient *cognitoidentityprovider.Client
)

func initIdentity(ctx context.Context) {
	idpOnce.Do(func() { idpClient = cognitoidentityprovider.NewFromConfig(loadAWSConfig(ctx)) })
}

// Once #2 — data-plane path. Mongo dial is EXPENSIVE; only the reconcile branch pays it.
var (
	dataOnce    sync.Once
	ebClient    *eventbridge.Client
	mongoClient *mongo.Client
)

func initDataPlane(ctx context.Context) {
	dataOnce.Do(func() {
		ebClient = eventbridge.NewFromConfig(loadAWSConfig(ctx))
		mongoClient = connectMongo(ctx) // the costly dial, deferred off the common path
	})
}

func handler(ctx context.Context, e MyEvent) (MyResp, error) {
	initIdentity(ctx) // common: only the cheap IdP client
	if needsReconcile(e) {
		initDataPlane(ctx) // rare: only this branch pays Mongo + EventBridge cold start
	}
	return doWork(ctx, e)
}

func main() { lambda.Start(handler) }

Anti-pattern to detect (greppable)

• A single func initAll() or one sync.Once that constructs every client.
• init() calling connectMongo(...), sql.Open(...), or any dial unconditionally.
• An eager package-level var mongoClient = mustConnect() outside any Once/branch.
• A sync.Once shared across two branches that never co-occur.

Decision aid

• Client used on (almost) every invocation, cheap to build? → fine to init eagerly or behind one Once.
• Client expensive (Mongo dial, slow-upstream TLS) and used on a minority branch? → its own sync.Once, called inside that branch only.
• On Provisioned Concurrency? → the asymmetry shrinks; still prefer segregation for the on-demand overflow path.

Related skills

• global-skills/aws-go/lambda-go-refactor-purge-audit/SKILL.md — when a branch (and its client) is removed, go mod tidy purges the module; audit the now-dead IAM actions.
• global-skills/aws-go/aws-sdk-go-v2-version-policy/SKILL.md — share one aws.Config; construct each chosen service client with NewFromConfig.

Sources

• Lambda best practices — Function code (init outside the handler, reuse the environment)
• sync.Once · aws-lambda-go/lambda (lambda.Start)
• Understanding and Remediating Cold Starts (AWS Compute Blog)

Last verified: 2026-06-03 against the AWS Lambda best-practices guide (live — "Initialize SDK clients and database connections outside of the function handler"), sync.Once semantics, and aws-lambda-go/lambda v1.54.0 (lambda.Start). Go has no SnapStart, confirmed against the SnapStart docs (Java/Python/.NET only). Re-check after: AWS SDK Go v2 major / CDK CLI major, or by 2026-09-03. Decay risk: low (the INIT-phase cost model and sync.Once are stable primitives). Found a drift? Run /skill-pattern-freshness-audit aws-go.

Source: esaldgut/ai-native-engineering-workspace — distributed by TomeVault.