AWS Lambda
AWS's serverless compute service — pay-per-invocation function execution with managed runtime, no server provisioning. **Now mounts S3 buckets as POSIX file systems via Amazon S3 Files** (May 2026), turning Lambda functions into ordinary file-I/O consumers of S3-resident data. Supports Python, Node.js, Java, Go, Ruby, .NET, custom runtimes. Default 15-minute max execution; 10 GB memory ceiling; 50 MB deployed package or 250 MB unzipped.
Definition
AWS's serverless compute service — pay-per-invocation function execution with managed runtime, no server provisioning. **Now mounts S3 buckets as POSIX file systems via Amazon S3 Files** (May 2026), turning Lambda functions into ordinary file-I/O consumers of S3-resident data. Supports Python, Node.js, Java, Go, Ruby, .NET, custom runtimes. Default 15-minute max execution; 10 GB memory ceiling; 50 MB deployed package or 250 MB unzipped.
Lambda eliminates the per-machine ops layer for event-driven and request/response compute — a function fires on an S3 PUT, an HTTP request, a SQS message, a CloudWatch event, etc. The 2026 S3 Files integration closes the long-standing gap where Lambda code couldn't read large S3-resident datasets without explicit `s3.get_object()` translation in every handler — now `open('/mnt/data/file.parquet')` works directly. Significant for serverless ML inference, file-conversion pipelines, on-demand log analysis, and any agentic-AI tooling that needs ephemeral compute over S3 data.
Event-driven processing on S3 PUTs, request/response API backends, on-demand serverless ML inference reading S3-resident model artifacts via S3 Files, file-conversion pipelines, scheduled jobs via EventBridge, agent-tool execution for AI orchestration platforms.
Recent developments
- Response Streaming now in all commercial AWS Regions (April 7, 2026). AWS extended Lambda's Response Streaming feature — sending response chunks back to the caller as they're produced — to global commercial-region availability. Critical for LLM-proxy workloads where streaming token output is the user-facing UX. Per AWS — Lambda Response Streaming Global Expansion.
- Container images up to 10 GB unzipped. ZIP packages remain capped at 50 MB compressed / 250 MB unzipped, but container images blow that cap up by 40× — PyTorch (800 MB), ffmpeg binaries, small LLMs, custom runtimes all fit. Per Viprasol — AWS Lambda Container Images in 2026 and AWS — Lambda Container Image Support announcement.
- 10 GB ephemeral /tmp storage. Lambda functions can write up to 10 GB of scratch state to /tmp during execution — supports model-loading, intermediate-data ETL, and "download → process → upload" patterns that previously had to chunk. Per AWS — Lambda 10 GB Ephemeral Storage.
- S3 Files mount integration (May 2026) lets Lambda treat S3 as POSIX. Mounting a bucket as NFS at /mnt —
open('/mnt/data/file.parquet')Just Works. Closes the long-standing "Lambda code can't easily read large S3-resident datasets" gap that forceds3.get_object()boilerplate in every handler. Per project notes + AWS Lambda — Create function with container image. - Durable Functions runtime support shipped. Lambda now exposes Durable Functions-style execution for orchestration workloads — replaces some of the Step Functions usage pattern with in-process saga/checkpoint primitives. Per AWS Lambda — Durable Functions supported runtimes.
- Custom runtimes + container approach is now the default for AI workloads. Community-recommended deployment pattern for Lambda+LLM is OCI-image-via-ECR-push, not ZIP-via-S3 — the 10 GB ceiling and the Docker familiarity outweigh ZIP's lower cold-start overhead for any non-trivial workload. Per Oneuptime — Deploy Lambda Functions as Container Images 2026.
Connections 5
Outbound 5
scoped_to1depends_on1consumes_via1enables1solves1