The APIs of higher level constructs in this module are experimental and under active development. They are subject to non-backward compatible changes or removal in any future version. These are not subject to the Semantic Versioning model and breaking changes will be announced in the release notes. This means that while you may use them, you may need to update your source code when upgrading to a newer version of this package.
Deploy Next.js apps on AWS with the AWS CDK.
- Supports all features of Next.js App and Pages Router for Node.js Runtime.
- Choose your AWS architecture for Next.js with the supported constructs:
NextjsGlobalFunctions,NextjsGlobalContainers,NextjsRegionalContainers,NextjsRegionalFunctions. - Global Content Delivery Network (CDN) built with Amazon CloudFront to deliver content with low latency and high transfer speeds.
- Serverless functions powered by AWS Lambda or serverless containers powered by AWS Fargate.
- Static assets (JS, CSS, public folder) are stored and served from Amazon Simple Storage Service (S3) for all constructs (except
NextjsRegionalContainers) to decrease latency and reduce compute costs by serving directly from S3. - Optimized images, data cache, full route cache and
'use cache: remote'are shared across compute with Amazon Simple Storage Service (S3) with supporting metadata in Amazon DynamoDB, andrevalidateTag/revalidatePathreach plain'use cache'on every instance. See'use cache'and'use cache: remote'. - Customize every construct via
overrides. - AWS security and operational best practices are utilized, guided by cdk-nag.
- First class support for monorepos.
- Bring Your Own Resources — import existing AWS resources (CloudFront distributions, ECS clusters, ALBs, S3 buckets, DynamoDB tables).
- AWS GovCloud (US) compatible with
NextjsRegionalFunctionsandNextjsRegionalContainers.
- Next.js app running v16.3 or higher (
next buildfails on older versions). If you don't have one yet - follow these steps to create one. - AWS Cloud Development Kit app either in the same package or separate package. cdk-nextjs supports monorepos.
- A Docker compatible container engine, but only for
NextjsGlobalContainersandNextjsRegionalContainers- we recommend Rancher Desktop with dockerd (moby). The two Functions types deploy zip Lambdas and need no container engine. - Node.js v24 (or LTS)
- Install
cdk-nextjsin the package(s) containing your CDK and Next.js app withnpm i cdk-nextjs - Deploy your Next.js app to AWS:
cdk deploy. Make sure you have AWS credentials configured. - Visit URL printed in terminal (CloudFormation Output) to view your Next.js app!
No next.config change is needed: cdk-nextjs runs next build itself, and sets NEXT_ADAPTER_PATH on that build so Next.js loads cdk-nextjs's Deployment Adapter.
Set adapterPath in next.config in three cases. An explicit adapterPath always wins over NEXT_ADAPTER_PATH, so this is also how you pin a specific adapter.
- cdk-nextjs isn't running your build —
skipBuild: true, or you build in CI and hand the output to CDK. - cdk-nextjs isn't a dependency of the Next.js app, only of the CDK app. cdk-nextjs resolves the adapter from your app's directory, so it can't find it in that layout.
- cdk-nextjs is linked from outside your project root — a
link:/file:dependency on a checkout elsewhere on disk, as this repo's own examples use. The symlink resolves out of the project, and the adapter derives its cache handler path from its own location, so Next.js gets acacheHandleroutsideturbopack.rootand Turbopack rejects it. SettingadapterPathfixes this because the build resolves it, after whatever put the adapter in place.
import { NextConfig } from "next";
const nextConfig: NextConfig = {
// ...
adapterPath: require.resolve("cdk-nextjs/adapter"),
};
export default nextConfig;require.resolve works in an ESM next.config.ts too, because Next.js transpiles the config to CJS before evaluating it.
For the Containers types, cdk-nextjs generates a Dockerfile in your Next.js app's directory. It cannot clean the file up after the image is built, so either commit it or gitignore it.
Every generated Dockerfile starts with the line # ~~ Generated by cdk-nextjs ~~, and cdk-nextjs overwrites any file carrying it — that way an upgrade that changes how the image is built does not leave a stale Dockerfile behind. To customize it, delete that first line; cdk-nextjs then leaves the file alone. If you edited a generated file without deleting the header, cdk-nextjs saves your version beside it as <name>.Dockerfile.bak before replacing it; gitignore that too.
import { App, Stack, StackProps } from "aws-cdk-lib";
import { Construct } from "constructs";
import { NextjsGlobalFunctions } from "cdk-nextjs";
import { join } from "node:path";
class WebStack extends Stack {
constructor(scope: Construct, id: string, props?: StackProps) {
super(scope, id, props);
new NextjsGlobalFunctions(this, "Nextjs", {
buildDirectory: join(import.meta.dirname),
});
}
}
const app = new App();
new WebStack(app, "web-stack");See examples/ for more usage examples.
The variables below are read at runtime by the cache handler inside your deployed
Lambda functions or containers. Set them on the compute through overrides:
// NextjsGlobalFunctions / NextjsRegionalFunctions
overrides: {
nextjsFunctions: {
functionProps: { environment: { CDK_NEXTJS_MEMORY_CACHE_TTL_MS: "0" } },
},
},
// NextjsGlobalContainers / NextjsRegionalContainers
overrides: {
nextjsContainers: {
taskImageOptions: { environment: { CDK_NEXTJS_MEMORY_CACHE_TTL_MS: "0" } },
},
},Values you pass are merged over the ones the construct sets, so the same mechanism overrides the infrastructure variables too.
Time to live in milliseconds for in-memory cache entries. After this duration, entries expire and are removed from the cache.
Note: Every memory hit is checked against the tag revalidation markers, so an entry revalidated on another instance is not served from memory. A tag's marker is read from DynamoDB only on its first use and a rolling re-read every few minutes; in between, revalidations arrive through the revalidation log (one Query per CDK_NEXTJS_TAG_REFRESH_MS). Memory still saves the S3 read and body parse. Set this to 0 to disable the memory cache entirely (all requests will fall through to S3 + DynamoDB).
Default: 3600000 (1 hour)
Examples:
- Short TTL (5 minutes):
300000- for frequently changing data - Medium TTL (1 hour):
3600000- balanced between freshness and performance - Long TTL (24 hours):
86400000- for mostly static content
Trade-offs:
- Lower values: More cache misses, fresher data, higher S3/DynamoDB costs
- Higher values: Fewer cache misses, better performance, but longer stale data windows
Maximum number of cache entries to store in memory. When this limit is reached, the least recently used (LRU) entry is evicted.
Default: 1000
Examples:
- Small cache:
100- minimal memory footprint for simple apps - Medium cache:
1000- good balance for typical applications - Large cache:
10000- for high-traffic apps with many unique pages
Trade-offs:
- Lower values: Less memory usage, more cache evictions
- Higher values: More memory usage, fewer cache evictions, better hit rates
Memory considerations: Each entry stores the full cache value (HTML, JSON, etc.). A typical page cache might be 10-100KB, so 1000 entries ≈ 10-100MB of memory. Consider your compute environment's memory limits (Lambda: 128MB-10GB, Fargate: 512MB-30GB) and size accordingly.
The adapter registers two cacheHandlers
next to the cacheHandler it already sets, unless your next.config names a
handler of its own for either:
| Directive | cacheHandlers name |
Stored | Tags |
|---|---|---|---|
'use cache' |
default |
each instance's memory | DynamoDB, all instances |
'use cache: remote' |
remote |
S3 cache bucket, memory in front | DynamoDB, all instances |
'use cache: remote'is shared: an entry any instance generates is the one every instance serves, until it expires or a tag is revalidated. Use it for results you don't want each Lambda instance or task to compute on its own. It costs an S3GetObjectthe first time an instance reads an entry, and aPutObjectper generated entry. Objects live under<buildId>/_use-cache/in the cache bucket and are pruned with the build.'use cache'stays per instance, as Next.js intends — butrevalidateTag,updateTagandrevalidatePathnow expire it on every instance, not only the one that ran them. Next.js's built-in handler keeps tags in process memory, so before this every other instance kept serving the revalidated value, even into pages re-rendered because of that revalidation.
Both read the same per-tag marker rows in the revalidation table that ISR and
the data cache use. A tag an instance has not seen yet — an entry read from S3,
a page's implicit revalidatePath tags — is read from its marker before it is
trusted, once per instance.
After that, an instance doesn't re-read the tags it tracks to learn what
changed. Each revalidateTag/updateTag also writes a row to a revalidation
log in the same table (pk = <buildId>#log, expiring after 15 minutes through
the table's ttl attribute). At most once per
CDK_NEXTJS_TAG_REFRESH_MS,
each instance sends one DynamoDB Query for the rows written since its last
one and applies those for tags it tracks. So a revalidation on another instance
is honored within that window (1 second by default), and on the instance that ran
it immediately. An idle instance's query costs 0.5 RCU a second, where
re-reading up to 1000 tracked tags cost about 500. Each tag's marker is still
re-read every 7.5 to 10 minutes, a few each second rather than all at once (about 8
RCU a second at the most an instance tracks, 10,000 tags). After a gap the log
may no longer cover (a Lambda frozen between invocations, a run of failed
queries), an instance forgets its tracked markers and reads each again as it's
needed. ISR and the data cache share the same copy of the markers, so one
revalidateTag writes each tag's rows once and one Query serves them all.
If you pass your own revalidationTable, enabling TTL on its ttl attribute is
recommended. Without it the log rows (about 100 bytes per revalidated tag) are
kept until a deploy replaces their build, whose post-deploy step deletes them.
Two instances that generate the same 'use cache: remote' entry at the same
moment both store it; each serves its own copy until that copy's revalidate
time, after which S3's is used.
How often, at most, in milliseconds, an instance asks the revalidation log what
other instances revalidated: the longest a revalidateTag elsewhere goes unseen,
for 'use cache', ISR and the data cache alike. It's one DynamoDB Query,
however many tags the instance tracks, shared by every request on the instance
during the window. 0 asks before every cache check.
Default: 1000
Size bound, in bytes, of each handler's in-memory store: all of 'use cache',
and the memory tier in front of S3 for 'use cache: remote'. Least recently used
entries are evicted first.
Default: 52428800 (50 MB, Next.js's own default)
The construct sets these itself and they typically don't need to be modified:
Unique identifier for the Next.js build. Used for cache isolation between deployments.
S3 bucket name for storing cached data (optimized images, data cache, full route cache, 'use cache: remote').
DynamoDB table name for tracking cache revalidations and tag-to-cache-key mappings.
Not set by default. Set it to cdk-nextjs:* to enable debug logs. This is especially useful to see cache handler activity.
Both are read out of your build, so you set them in one place — your app — and the infrastructure follows.
Becomes the S3 key prefix for static assets and the prefix on every CloudFront
cache behavior. See Resource Isolation; the basePath
prop only exists to be explicit about it.
- A path (
assetPrefix: "/cdn") — Next.js emits bundle URLs as/cdn/_next/static/..., on top ofbasePathrather than under it, while the objects stay in S3 under<basePath>/_next/static/....NextjsGlobalFunctionsandNextjsGlobalContainersadd a cache behavior for<assetPrefix>/_next/static*pointing at the assets bucket, with a CloudFront Function that rewrites the prefix back to the S3 keys. It costs one of the 25 cache behaviors cdk-nextjs budgets for you. Not supported onNextjsRegionalFunctionsorNextjsRegionalContainers— nothing there maps the prefix back, so bundles would 404; synth warns if your app sets one. - An absolute URL (
assetPrefix: "https://cdn.example.com") — names an origin cdk-nextjs doesn't serve, so nothing is added, on any construct. Point that host at the assets bucket yourself, including thebasePathprefix the objects are stored under. - An absolute URL with a path (
assetPrefix: "https://cdn.example.com/cdn") — its path is treated exactly like the path case above, becausenext buildcompiles a/cdn/_next/:path+rewrite of its own:next startserves bundles under that path too, so the Global constructs answer there as well. Useful when that host is a second domain on the same distribution.
Only for NextjsGlobalFunctions and NextjsRegionalFunctions, and only if you
need it. AWS Lambda caps a function's unzipped code at 250 MB. If your app
exceeds it, cdk-nextjs fails at synth with the measured size and points you here:
Function group "default" is 274 MB unzipped, over Lambda's 250 MB limit. Use the
`functionGroups` prop to package routes into separate functions. Note that
splitting only removes route-local code — anything reachable from a shared layout
or the `next` runtime is in every group.
functionGroups declares which routes get their own Lambda. Everything you do
not name stays on the default function:
new NextjsGlobalFunctions(this, "Nextjs", {
buildDirectory: join(import.meta.dirname, "..", "web"),
functionGroups: [
{ name: "reports", routes: ["/dashboard/reports/**"] },
{ name: "admin", routes: ["/admin", "/admin/**", "/settings"] },
],
});Each group becomes one Lambda function containing only the routes it owns (plus
the framework closure every function needs), and one CloudFront behavior — or API
Gateway resource, for NextjsRegionalFunctions — per pattern.
Route patterns are either an exact path (/settings) or a subtree
(/admin/**). A subtree owns what is under it, not the path itself:
/admin/** does not claim /admin. To own a page and everything under it, list
both, as the example above does: ["/admin", "/admin/**"]. Where two groups
could both match, the longest pattern wins, so /api/** and /api/reports/**
can coexist in different groups. Dynamic segments (/blog/[slug]), route group
segments (/(marketing)/about), /, and /** are rejected at synth —
CloudFront matches literal path prefixes and cannot express them. /index is
accepted for an App Router app/index/page.tsx, but not for a Pages Router home
page, which is the same file as /.
What is routed for you. Patterns decide which files a group packages; the edge then has to send every URL those files serve to that group. cdk-nextjs derives these without extra patterns:
- A Pages Router page's data URL,
/_next/data/<buildId>/<page>.json. The behavior carries the literal build ID, so it changes on every deploy; a client still on the previous build falls through to thedefaultfunction, which answers 404, and the Next.js client reloads the page. - The
trailingSlashform of an exact pattern (/pricing/). CloudFront gets a behavior for it; onNextjsRegionalFunctionsnone is needed, because API Gateway already routes/pricing/to the/pricingresource. - The parent of an optional catch-all.
/shop/**movingapp/shop/[[...slug]]/page.tsxalso routes/shop, which the same file serves. Do not add/shopyourself: it matches no route and is rejected. - An interception route
is packaged with the group that owns the URL it intercepts, not by its own
path:
app/feed/(..)photo/[id]follows/photo/[id], because a soft navigation requests/photo/1and Next.js rewrites it in whichever function receives it.
What synth rejects. After assigning routes, cdk-nextjs replays the CloudFront behaviors over every URL each file serves and fails when one would reach a function without its file, naming the file, the URL and the pattern to add. That covers:
- A file behind several templates that a pattern only partly covers — root
params (
app/[locale]/page.tsxprerendered as/enand/de) grouped with/enalone. A dynamic first segment can never be routed, so such a file stays indefault. - An intercepted URL space split between groups (
/photo/[id]in one,/photo/1in another), since the intercepting file cannot be in both. - A
next.configrewrite whose source reaches one group and whose destination file is in another: rewrites run inside the function that received the request. Only rewrites with a literal destination are checked. - A group pattern that duplicates a top-level
public/entry's behavior (public/docs/and/docs/**), or sits under one (/docs/guide/**), which would send the group's routes to S3.
Not checked, because it is only known at request time: a middleware
NextResponse.rewrite() to a route in another group reaches the function that
received the original URL, which lacks the route's file. It answers 404 (or a
308 to the canonical URL when only the case differs) and logs a warning naming
the group that owns the route. Keep a middleware rewrite's source and
destination in the same group. The same goes
for a next.config rewrite whose destination is built from its parameters
(/b/:slug → /blog/:slug), and for a dynamic segment of a default route
that overlaps a group's literal path (/[section]/intro against /docs/**).
Likewise a Pages Router res.revalidate() renders the page in the function
that called it, so it can only revalidate pages in its own group; the error
names the group that owns the page. Use revalidatePath() or revalidateTag()
across groups. And an <Image> whose src is a route in another group
(/api/avatar/42 under an /api/** group) can't be optimized: /_next/image
runs in the default function, which fetches a local source that isn't a
public/ or _next/static file by running its route in-process, and it doesn't
carry that route, so the image request fails. Keep image-serving routes in
default, or point src at a public/ file or a remote URL.
What splitting does and does not save. Every function ships the same next
runtime closure, so splitting only moves route-local code and its dependencies.
A group whose routes import a large library is worth extracting; splitting an app
in half does not halve either function.
Notes:
- Group names must be valid CDK construct ids (letters, digits,
-).defaultis reserved. - A pattern that matches no route in the build is a synth error, not a silent no-op — it is almost always a typo.
- Cannot be combined with
i18n: locale-prefixed routes would need one CloudFront behavior per locale per pattern. - Group patterns (and the routes derived from them above) count against the
same CloudFront behavior budget as your
public/entries (see Limitations). - Per-group
overrideslet you size each function independently:{ name: "reports", routes: [...], overrides: { functionProps: { memorySize: 2048 } } }.
How requests are served inside the compute — the adapter runtime, and where it deliberately differs from next start — is in docs/adapter-runtime.md.
Architecture includes AWS Lambda Functions to respond to dynamic requests and CloudFront Distribution to globally serve requests and distribute static assets. Use this construct when you have unpredictable traffic, can afford occasional latency (i.e. cold starts - typically 1% of production traffic), and/or want the most granular pricing model. (code)
architecture-beta
group aws(cloud)[AWS Cloud]
group cache[Cache Layer] in aws
service user(internet)[User]
service cloudfront(server)[CloudFront Distribution] in aws
service s3static(disk)[S3 Static Assets] in aws
service lambda(server)[Lambda Function] in aws
service dynamodb(database)[DynamoDB Table] in cache
service s3cache(disk)[S3 Cache Bucket] in cache
user:R --> L:cloudfront
cloudfront:R --> L:s3static
cloudfront:R --> L:lambda
lambda:R --> L:dynamodb
lambda:R --> L:s3cache
Architecture includes ECS Fargate containers to respond to dynamic requests and CloudFront Distribution to globally serve requests and distribute static assets. Use this option when you have predictable traffic, need the lowest latency, and/or can afford a less granular pricing model. (code)
architecture-beta
group aws(cloud)[AWS Cloud]
group vpc[VPC] in aws
group cache[Cache Layer] in aws
service user(internet)[User]
service cloudfront(server)[CloudFront Distribution] in aws
service s3static(disk)[S3 Static Assets] in aws
service alb(server)[Application Load Balancer] in vpc
service fargate(server)[ECS Fargate Containers] in vpc
service dynamodb(database)[DynamoDB Table] in cache
service s3cache(disk)[S3 Cache Bucket] in cache
user:R --> L:cloudfront
cloudfront:R --> L:s3static
cloudfront:R --> L:alb
alb:R --> L:fargate
fargate:R --> L:dynamodb
fargate:R --> L:s3cache
Architecture includes ECS Fargate containers to respond to dynamic requests and Application Load Balancer to regionally serve requests. Use this options when you cannot use Amazon CloudFront (i.e. AWS GovCloud). (code)
architecture-beta
group aws(cloud)[AWS Cloud]
group vpc[VPC] in aws
group cache[Cache Layer] in aws
service user(internet)[User]
service alb(server)[Application Load Balancer] in vpc
service fargate(server)[ECS Fargate Containers] in vpc
service dynamodb(database)[DynamoDB Table] in cache
service s3cache(disk)[S3 Cache Bucket] in cache
user:R --> L:alb
alb:R --> L:fargate
fargate:R --> L:dynamodb
fargate:R --> L:s3cache
Architecture includes AWS Lambda Functions to respond to dynamic requests and API Gateway REST API to regionally serve requests and distribute static assets. Use this options when you cannot use Amazon CloudFront (i.e. AWS GovCloud). (code)
architecture-beta
group aws(cloud)[AWS Cloud]
group cache[Cache Layer] in aws
service user(internet)[User]
service apigateway(server)[API Gateway REST API] in aws
service s3static(disk)[S3 Static Assets] in aws
service lambda(server)[Lambda Function] in aws
service dynamodb(database)[DynamoDB Table] in cache
service s3cache(disk)[S3 Cache Bucket] in cache
user:R --> L:apigateway
apigateway:R --> L:s3static
apigateway:R --> L:lambda
lambda:R --> L:dynamodb
lambda:R --> L:s3cache
The simplest path to deploy Next.js is on Vercel - the Platform-as-a-Service company behind Next.js. However, deploying to Vercel can be expensive and some developers want all of their workloads running directly on AWS. Developers can deploy Next.js on AWS through AWS Amplify Hosting, but Amplify does not support all Next.js features and manages AWS resources for you so they cannot be customized. If Amplify meets your requirements we recommend you use it, but if you want to use all Next.js features or want more visibility into the AWS resources then this construct is for you.
- Integrate through Next.js's official interfaces. Build on the Deployment Adapter API and
@next/routingrather than Next.js internals, so new Next.js versions rarely break this construct. - Security first.
- One architecture does not fit all.
- Enable customization everywhere.
cdk-nextjs supports importing existing AWS resources instead of creating new ones. This is especially useful for per-branch (MR/PR) environments where you deploy shared infrastructure (VPC, ECS cluster, ALB) once and spin up lightweight branch stacks that reuse it.
| Resource | Prop | Available On |
|---|---|---|
| S3 Cache Bucket | cacheBucket |
All constructs |
| DynamoDB Revalidation Table | revalidationTable |
All constructs |
| S3 Static Assets Bucket | staticAssetsBucket |
All constructs |
| CloudFront Distribution | distribution |
NextjsGlobalFunctions, NextjsGlobalContainers |
| ECS Cluster | ecsCluster |
NextjsGlobalContainers, NextjsRegionalContainers |
| ALB | alb |
NextjsGlobalContainers, NextjsRegionalContainers |
- Cache bucket and DynamoDB table must be dedicated to one deployment. Sharing either between deployments (branches, stages, apps) is not supported: each deploy's post-deploy step prunes every cache object and revalidation entry that isn't from its own current build, so deploying one branch wipes the others' caches. Buckets and tables are cheap, so give each deployment its own, or leave
cacheBucket/revalidationTableunset and let cdk-nextjs create them (they are removed with the stack). - Static assets bucket can be shared, but only between deployments with different
basePaths, which become their S3 key prefixes. Pruning only looks at<basePath>/_next/, so it never touches another prefix. Two deployments under the same prefix (for example two subdomain previews both at the root) overwrite each other'spublic/files and prune each other's_next/assets once they are past the prune TTL (30 days by default), so give those their own buckets.NextjsGlobalFunctionsandNextjsGlobalContainersread thebasePathout of your Next.js build and use it as the S3 key prefix, since CloudFront serves assets from S3 by request path and the two can't differ. Set thebasePathprop only if you want to be explicit about it; a value that disagrees with your app's fails at synth.
ApplicationLoadBalancedFargateService always creates a listener on port 80 — there is no opt-out. When you import an ALB that already has a listener on that port, the duplicate causes a deployment failure. Since CDK doesn't expose a way to prevent this, removeAutoCreatedListener() surgically removes the generated CloudFormation resources: the CfnListener, its security group ingress rule, rebuilds the ECS service DependsOn without the deleted listener, and removes auto-created CfnOutput resources:
const nextjs = new NextjsRegionalContainers(this, "Nextjs", {
// ...
alb: sharedAlb,
ecsCluster: sharedCluster,
});
nextjs.nextjsContainers.removeAutoCreatedListener();See examples/bring-your-own/ for a complete deployable example with a shared VPC, ECS cluster, and ALB, per-branch host-header routing, and a cache bucket, revalidation table, and static assets bucket per branch.
cdk-nextjs can deploy ephemeral preview environments per merge request (MR) or pull request (PR). The recommended approach uses subdomain-based routing (pr-123.app.example.com) so each preview environment runs the same Next.js build as production — no basePath configuration or separate builds required.
See examples/bring-your-own/ for a fully deployable example using NextjsRegionalContainers with a shared VPC, ALB, and ECS Cluster (connected via SSM Parameter Store), and S3 buckets and a DynamoDB table owned by each branch stack.
Each preview environment needs its own cache bucket and revalidation table (see Resource Isolation). Since every preview serves the same build at the root, give each its own static assets bucket too. Leaving cacheBucket, revalidationTable, and staticAssetsBucket unset does this: cdk-nextjs creates them in the branch stack and deletes them when it's torn down.
- Wildcard DNS record:
*.app.example.com→ your routing layer (ALB, CloudFront, API Gateway) - Wildcard ACM certificate:
*.app.example.com
Subdomain routing via ALB host-based listener rules. This is the fastest and simplest approach. For NextjsGlobalContainers, CloudFront forwards the Host header to the ALB origin, so the ALB handles all branch routing — no CloudFront changes needed per branch.
See examples/bring-your-own/ for the full implementation.
Subdomain routing via API Gateway custom domain mappings.
- Per branch, deploy a cdk-nextjs stack that creates its own API Gateway, Lambda function, S3 buckets, and DynamoDB table
- Create an API Gateway custom domain (
pr-123.app.example.com) mapped to the branch's API stage - Tear down the branch stack on MR close
Requires a separate cdk-nextjs stack per branch, each with its own CloudFront distribution. CloudFront cannot route to different Lambda Function URL origins based on the Host header — origin selection is determined by cache behavior path patterns, not request headers.
- Per branch, deploy a full cdk-nextjs stack that creates its own CloudFront distribution, Lambda function, S3 buckets, and DynamoDB table
- Point
pr-123.app.example.comDNS to the branch's CloudFront distribution - Tear down the branch stack on MR close
Note: CloudFront distributions take several minutes to create/update, so this architecture has the slowest preview environment spin-up time.
- If using
NextjsGlobalFunctionsorNextjsGlobalContainers(which use CloudFront), each top level file/directory inpublic/takes a CloudFront cache behavior, and a distribution's default quota is 75 behaviors (cdk-nextjs uses 3 to 6 itself). We recommend you put all of your public assets into one top level directory (i.e. public/static) so you don't reach this limit. If you raise the quota for your account, setmaxCacheBehaviorson the distribution props (overrides.nextjsGlobalFunctions.nextjsDistributionProps, or the Containers equivalent) to match. The synth check only counts the behaviors cdk-nextjs adds, so if you pass your owndistribution(or add behaviors throughoverrides.nextjsDistribution.distributionProps), lowermaxCacheBehaviorsby the behaviors it already has. See CloudFront Quotas for more information. - If using
NextjsGlobalFunctionsorNextjsGlobalContainers, on-demand revalidation (revalidatePath, revalidateTag) creates the CloudFront invalidations for you (so does the Pages Router'sres.revalidate(), for the page's HTML and_next/datacopies), but invalidations are eventually consistent: until one completes (usually seconds, occasionally minutes), some edge locations can still serve the previous copy.- Burst limit. CloudFront allows 15 wildcard invalidation paths in progress per distribution, and each page cdk-nextjs invalidates is one wildcard path (
/blog*). Next.js sends everyrevalidatePath/revalidateTagfrom one request as a single call, one invalidation per cache-life profile, however many pages it covers. A Pages Routerres.revalidate()sends one invalidation per call, and a page regenerated after a stale-while-revalidaterevalidateTag(tag, profile)sends one more when it lands. When the quota is full, cdk-nextjs retries the request as a single/*(a colder edge, never a stale page). If that is also rejected, it logsFailed to create CloudFront invalidationand the page stays stale at the edge until itss-maxageexpires. - For bulk updates (a CMS webhook touching many pages), call
revalidateTagwith a tag the pages share, orrevalidatePathfor each page within one request, rather than loopingres.revalidate(). If you need to know when invalidations are dropped, add a CloudWatch Logs metric filter on that log line.
- Burst limit. CloudFront allows 15 wildcard invalidation paths in progress per distribution, and each page cdk-nextjs invalidates is one wildcard path (
- If using
NextjsGlobalFunctions, a client-suppliedAuthorizationheader does not reach your app: CloudFront signs each request to the Lambda Function URL with SigV4, which uses that header. Send credentials under a different name (e.g.x-authorization) and read that header in your app. cdk-nextjs does not rename it for you — it no longer uses AWS Lambda Web Adapter, soAWS_LWA_AUTHORIZATION_SOURCEno longer applies. - If using
NextjsGlobalFunctions, aPOSTorPUTwith a request body must carry anx-amz-content-sha256header holding the hex SHA-256 of that body. This is AWS behavior, not a cdk-nextjs choice: CloudFront signs the origin request to the Lambda Function URL with SigV4 but does not hash the body, and Lambda rejects an unsigned payload with403 InvalidSignatureException. Requests from your own pages are handled for you — cdk-nextjs injects afetch/XMLHttpRequestwrapper into the client bundle that adds the header, so server actions, route handlers,fetchandXMLHttpRequestfrom the browser work untouched, whatever the body type. The exception is a native<form method="post">submitted before the page hydrates (a server action form without JavaScript, or clicked before the bundle loads): the browser sends it, not the wrapper, so it has no header and gets the 403. Anything that is not that browser bundle has to add the header itself:curl, a mobile app, a webhook provider, another service calling your API routes. Server-to-server callers can compute it with the AWS SDK'sSha256, or withcrypto.createHash("sha256").update(body).digest("hex"). The other three deployment types are unaffected. - If using
NextjsGlobalFunctions, aHEADrequest is answered withcontent-length: 0, whatever length your handler declares. This is Lambda Function URL behavior: the same runtime's declared length reaches the client through API Gateway (NextjsRegionalFunctions) and from an ALB, with or without CloudFront in front (the two Containers types). - Group patterns declared via
functionGroupsalso count against the CloudFront behavior limit above. - If using
NextjsRegionalFunctionswithout a custom domain, API Gateway REST APIs require a stage name (default:/prod) to be specified. This causes links to pages and static assets to break because they're not prefixed with the stage name. You can work around this issue by specifying basePath in next.config.js as your stage name — leave the construct'sbasePathprop unset when you do, since API Gateway strips the stage before matching resources. API Gateway strips the stage from the path it passes to Lambda; cdk-nextjs puts it back for an app whosebasePathstarts with it, reading the stage off each request, so no middleware is needed and a renamed stage needs no configuration. Mapping a custom domain at the root avoids all of this — see examples/regional-functions/README.md.
This construct by default implements all AWS security best practices that a CDK construct library reasonably can considering cost and complexity. Below are additional security practices we recommend you implement within your CDK app. Please see them below:
- VPC Flow Logs. See examples/ for sample implementation.
- Scan ECR Images For Vulnerabilities.
- For
NextjsGlobalFunctionsandNextjsGlobalContainers, CloudFront Access Logs. See examples/ for sample implementation. - For
NextjsGlobalContainersandNextjsRegionalContainers, use ALB HTTPS Listener - If using
NextjsGlobalContainersandNextjsRegionalContainers, enableReadonlyRootFilesystem. This will remove ability to use Static On-Demand feature of Next.js so it's not enabled by default, but is recommended for security.
The following basic assumptions were used for a typical medium Next.js app. See docs/usage.xlsx for detailed assumptions and usage per construct type that you can plug into AWS Pricing Calculator.
| Metric | Value |
|---|---|
| Monthly Active Users | 1K |
| Pages Visited Per Month Per User | 100 |
| Avg Request Size | 50KB |
| Static Requests Per Page (js, css, etc) | 15 |
| Static Requests Cache Hit % | 50% |
| Static Assets Size | 10GB |
| Dynamic Requests Per Page (document, optimized images, etc.) | 5 |
| Dynamic Cache Read % | 50% |
| Dynamic Cache Write % | 5% |
| Dynamic Cache Data Size | 10GB |
| Average Dynamic Cache Request Size | 100KB |
| Dynamic Revalidate Tag % | 1% |
More Details:
- Assume ARM architecture for compute
- AWS Region: us-east-1
- Excludes charges related to: CloudWatch Logs, NAT Gateway data processing
| Service | Monthly Usage | Estimated Monthly Cost (USD) |
|---|---|---|
| Lambda | 500K requests, 2 GB memory, 150ms avg duration | $0.00 (Always Free Tier) |
| CloudFront | 2M requests, 100 GB transfer to internet | $0.00 (Always Free Tier) |
| S3 | 20 GB storage, 1M GET, 250K PUT requests | $2.11 |
| DynamoDB | 10 GB storage, 5K Reads, 5K Writes | $2.50 |
| Total | $8.09 |
| Service | Monthly Usage | Estimated Monthly Cost (USD) |
|---|---|---|
| ECS Fargate | 1 task (1 vCPU, 2 GB) | $28.44 |
| ALB | 1 LB, 1.04GB/hr, 5.79 conn/sec | $22.50 |
| CloudFront | 2M requests, 100 GB transfer to internet | $0.00 (Always Free Tier) |
| S3 | 20 GB storage, 1M GET, 250K PUT requests | $2.11 |
| DynamoDB | 10 GB storage, 5K Reads, 5K Writes | $2.50 |
| VPC | NAT Gateway, 2 AZs | $65.70 |
| Total | $121.25 |
| Service | Monthly Usage | Estimated Monthly Cost (USD) |
|---|---|---|
| ECS Fargate | 1 task (2 vCPU, 4 GB), always on | $28.44 |
| ALB | 1 LB, 4.17 GB/hr, 23.15 conn/sec | $40.78 |
| S3 | 10 GB storage, 250K GET, 250K PUT requests | $1.58 |
| DynamoDB | 10 GB storage, 5K Reads, 5K Writes | $2.50 |
| VPC | NAT Gateway, 2 AZs | $65.70 |
| Total | $139.00 |
| Service | Monthly Usage | Estimated Monthly Cost (USD) |
|---|---|---|
| Lambda | 500K requests, 2 GB memory, 150ms avg duration | $0.00 (Always Free Tier) |
| API Gateway | 2M requests | $7.00 |
| S3 | 10 GB storage, 250K GET, 250K PUT requests | $1.58 |
| DynamoDB | 10 GB storage, 5K Reads, 5K Writes | $2.50 |
| Total | $11.08 |
Measured with k6 from EC2 in us-east-1, the region of the stacks under test (one load generator per stack), against a small benchmark app (examples/bench-app) that makes no calls outside its own stack. Functions run the default 2048 MB Lambda functions. Containers run the default 1 vCPU / 2 GB Fargate tasks, with CPU autoscaling from 2 to 10 tasks added (the default is one task and no autoscaling). Methodology, caveats, and how to reproduce the numbers: examples/load-tests.
What each route measures:
static-asset,static,image: CloudFront's cache on the Global constructs, and the construct itself on the Regional ones.isr: the cache until the page is stale (10 s), then a background revalidation through the S3 + DynamoDB cache.ssr,stream,rsc,api: the construct's compute on every request, on every construct.
| Construct | ssr at 10 req/s |
ssr at 50 req/s |
ssr capacity |
isr capacity |
Cold start p50 |
|---|---|---|---|---|---|
NextjsGlobalFunctions |
36 / 103 | 33 / 84 | ≥ 6560 req/s | ≥ 6560 req/s | 1390 ms |
NextjsGlobalContainers |
19 / 44 | 18 / 219 | 194 req/s | ≥ 6400 req/s | – |
NextjsRegionalContainers |
22 / 172 | 1972 / 39536 |
675 req/s | ≥ 6075 req/s | – |
NextjsRegionalFunctions |
32 / 114 | 30 / 103 | ≥ 6560 req/s | ≥ 6560 req/s | 1380 ms |
Latency is p50 / p99 in ms, with every route loaded at that rate at once. Capacity is the highest rate sustained with p99 under 1000 ms and under 1% errors.
What this means for choosing one:
- Functions constructs scale with no sizing. Both held at least 6,560 req/s of SSR, the load generator's limit, at the same latency as at 10 req/s. The cost is a cold start of about 1.4 s on a new execution environment.
- Containers constructs are fastest per request but need sizing. Two 1 vCPU tasks served SSR at 19–22 ms p50 against Lambda's ~30 ms, but tasks take minutes to add. Past capacity they don't degrade gradually: tasks too busy to answer the ALB health check are replaced, which leaves fewer tasks for the same load. Leave headroom.
- CloudFront carries the Global constructs' cached traffic. Prerendered pages, ISR hits, images and assets were answered at the edge in 2–3 ms at every rate. On the Regional constructs every request reaches the construct:
NextjsRegionalContainersserves assets and image optimization from the same tasks, so 50 req/s on every route at once (400 req/s) is already more than two tasks handle. - Cached pages scale on every construct. ISR hits held at least 6,000 req/s on all four: at the edge on the Global constructs, and through the S3 + DynamoDB cache on the Regional ones, where each instance holds a tag's revalidation marker for a second (
CDK_NEXTJS_TAG_REFRESH_MS, see the Caching Guide).
NextjsGlobalFunctions
Latency per route, p50 / p99 in ms, with every route at the same rate at once:
| Route | 10 req/s | 50 req/s | 250 req/s | 1000 req/s | CDN hits |
|---|---|---|---|---|---|
| static-asset | 2.1 / 4.8 | 2.1 / 4.7 | 2.0 / 4.9 | 2.1 / 11 | 100% |
| static | 3.3 / 7.5 | 3.2 / 6.8 | 3.1 / 7.3 | 3.2 / 13 | 100% |
| isr | 3.3 / 7.4 | 3.3 / 7.7 | 3.2 / 7.2 | 3.2 / 12 | 100% |
| ssr | 36 / 103 | 33 / 84 | 31 / 86 | 29 / 73 | 0% |
| stream | 233 / 268 | 230 / 261 | 228 / 259 | 226 / 252 | 0% |
| stream (TTFB) | 32 / 89 | 29 / 66 | 27 / 65 | 25 / 57 | |
| rsc | 32 / 67 | 28 / 61 | 27 / 60 | 25 / 52 | 0% |
| api | 29 / 79 | 26 / 55 | 24 / 61 | 23 / 47 | 0% |
| image | 3.2 / 6.3 | 3.1 / 6.8 | 3.1 / 7.0 | 3.1 / 12 | 100% |
Capacity: isr ≥ 6560 req/s, ssr ≥ 6560 req/s.
Cold start, time to first byte of /ssr on a new execution environment: p50 1390 ms, p90 1451 ms (warm: 84 ms).
Browser, p75 in ms, under 50 req/s per route:
| Route | LCP | TTFB | Client-side navigation |
|---|---|---|---|
| static | 76 | 13 | 15 |
| ssr | 80 | 39 | 53 |
| isr | 52 | 8.7 | 12 |
| stream | 64 | 34 | 38 |
| image | 52 | 6.7 | 10 |
NextjsGlobalContainers
Latency per route, p50 / p99 in ms, with every route at the same rate at once:
| Route | 10 req/s | 50 req/s | CDN hits |
|---|---|---|---|
| static-asset | 1.4 / 3.9 | 1.5 / 3.7 | 100% |
| static | 1.9 / 4.8 | 1.8 / 3.7 | 100% |
| isr | 2.0 / 4.4 | 1.7 / 4.0 | 100% |
| ssr | 19 / 44 | 18 / 219 | 0% |
| stream | 217 / 242 | 217 / 359 | 0% |
| stream (TTFB) | 17 / 45 | 16 / 207 | |
| rsc | 12 / 26 | 12 / 86 | 0% |
| api | 11 / 23 | 10 / 81 | 0% |
| image | 2.0 / 4.7 | 1.7 / 4.1 | 100% |
Capacity: isr ≥ 6400 req/s, ssr 194 req/s.
Browser, p75 in ms, under 50 req/s per route:
| Route | LCP | TTFB | Client-side navigation |
|---|---|---|---|
| static | 88 | 12 | 17 |
| ssr | 80 | 26 | 39 |
| isr | 56 | 6.1 | 13 |
| stream | 64 | 20 | 25 |
| image | 52 | 5.4 | 11 |
NextjsRegionalContainers
Latency per route, p50 / p99 in ms, with every route at the same rate at once:
| Route | 10 req/s | 50 req/s | CDN hits |
|---|---|---|---|
| static-asset | 18 / 159 | 1698 / 28824 |
no CDN |
| static | 41 / 163 | 916 / 28298 |
no CDN |
| isr | 40 / 169 | 697 / 27146 |
no CDN |
| ssr | 22 / 172 | 1972 / 39536 |
no CDN |
| stream | 221 / 354 | 1909 / 27423 |
no CDN |
| stream (TTFB) | 19 / 166 | 1881 / 27162 | |
| rsc | 9.4 / 64 | 433 / 26066 |
no CDN |
| api | 6.9 / 55 | 429 / 26041 |
no CDN |
| image | 140 / 440 | 1599 / 27489 |
no CDN |
Capacity: isr ≥ 6075 req/s, ssr 675 req/s.
Browser, p75 in ms, under 10 req/s per route. static is each visit's first page: on a plain-HTTP origin, Chrome tries HTTPS first and falls back after about 3 s (HTTPS-Upgrades), which real visitors pay too. Put a certificate on the ALB to avoid it.
| Route | LCP | TTFB | Client-side navigation |
|---|---|---|---|
| static | 3100 | 3047 | 16 |
| ssr | 68 | 14 | 27 |
| isr | 88 | 43 | 13 |
| stream | 48 | 11 | 13 |
| image | 212 | 38 | 9.6 |
NextjsRegionalFunctions
Latency per route, p50 / p99 in ms, with every route at the same rate at once:
| Route | 10 req/s | 50 req/s | 250 req/s | 1000 req/s | CDN hits |
|---|---|---|---|---|---|
| static-asset | 34 / 76 | 32 / 72 | 30 / 72 | 30 / 68 | no CDN |
| static | 58 / 138 | 57 / 140 | 57 / 147 | 57 / 127 | no CDN |
| isr | 58 / 149 | 56 / 124 | 56 / 145 | 55 / 125 | no CDN |
| ssr | 32 / 114 | 30 / 103 | 29 / 115 | 29 / 98 | no CDN |
| stream | 229 / 293 | 227 / 288 | 227 / 299 | 226 / 282 | no CDN |
| stream (TTFB) | 27 / 85 | 25 / 84 | 25 / 97 | 24 / 76 | |
| rsc | 27 / 76 | 26 / 73 | 25 / 83 | 25 / 70 | no CDN |
| api | 24 / 84 | 22 / 70 | 22 / 97 | 22 / 67 | no CDN |
| image | 127 / 210 | 125 / 221 | 125 / 236 | 123 / 211 | no CDN |
Capacity: isr ≥ 6560 req/s, ssr ≥ 6560 req/s.
Cold start, time to first byte of /ssr on a new execution environment: p50 1380 ms, p90 1442 ms (warm: 71 ms).
Browser, p75 in ms, under 50 req/s per route:
| Route | LCP | TTFB | Client-side navigation |
|---|---|---|---|
| static | 128 | 70 | 17 |
| ssr | 88 | 35 | 50 |
| isr | 108 | 62 | 13 |
| stream | 64 | 31 | 36 |
| image | 220 | 60 | 9.6 |
Steps to build locally:
git clone https://github.com/cdklabs/cdk-nextjs.gitcd cdk-nextjspnpm i && pnpm compile && pnpm build
This project uses Projen, so make sure to not edit Projen created files and only edit .projenrc.ts.
Q: How does this compare to cdk-nextjs-standalone or OpenNext?
A: cdk-nextjs builds on Next.js's official Deployment Adapter API and runs requests through @next/routing, the routing library Next.js publishes for platform adapters. Next.js tells cdk-nextjs what it built, and routing follows the same rules as next start. Nothing re-implements Next.js's router or patches its server. The few Next.js internals cdk-nextjs does load (the image optimizer and static file server) are loaded as they are, not modified. This keeps upgrades cheap and breaking changes rare. cdk-nextjs is tested against Next.js's own end-to-end test suite; harness-coverage.md lists the cases that differ and why.
Q: Why does cdk-nextjs depend upon Next.js v16.3 or higher?
A: cdk-nextjs builds through the public Adapter API, and 16.3 is the first version whose build output carries everything it reads (middleware matchers and per-file asset hashes). 16.2 also added the Image Optimization Caching it relies on. The runtime bundles @next/routing at the version it was tested against (see package.json), and newer Next.js 16 releases are expected to work.
Q: How does cdk-nextjs support caching in Next.js?
A: See Caching Guide
Q: How customizable is the cdk-nextjs package for different use cases?
A: The cdk-nextjs package offers deep customization through prop-based overrides. These can be accessed in the construct props, allowing you to override settings like VPC configurations, CloudFront distribution, and ECS/Fargate setup. For example, you can modify nextjsBuildProps to customize the build process or use nextjsDistributionProps to adjust how CloudFront handles caching and routing. This level of control makes it easy to adapt the infrastructure to your application’s specific performance, networking, or deployment needs.
Q: How can I use a custom domain with cdk-nextjs?
A: See low-cost example.
Q: What is difference between NextjsGlobalFunctionsProps.overrides.nextjsDistribution and NextjsGlobalFunctionsProps.overrides.nextjsGlobalFunctions.nextjsDistributionProps
A: NextjsGlobalFunctionsProps.overrides.nextjsDistribution allows you to customize any construct's props within NextjsDistribution and is likely what you want whereas NextjsGlobalFunctionsProps.overrides.nextjsGlobalFunctions.nextjsDistributionProps allows you to customize the props passed into the construct: NextjsDistribution. This principle also applies to other similarly named overrides.
Q: How can I cdk bootstrap --cloudformation-execution-policies ... my AWS Account with limited permissions for cdk-nextjs to deploy?
A: See docs/cdk-nextjs-cfn-exec-policy.json. Note, this IAM Policy is scoped to all cdk-nextjs constructs so you can remove services if you know the construct you're using doesn't use that service.
This construct was built on the shoulders of giants. Thank you to the contributors of cdk-nextjs-standalone and OpenNext.
Thank you for helping other developers deploy Next.js apps on AWS