SYSTEM Cited by 2 sources
AWS Lambda SnapStart¶
Lambda SnapStart (launched 2022) is AWS Lambda's snapshot-based cold-start acceleration, built on Firecracker VM snapshotting. Instead of initializing each function invocation's execution environment from scratch, SnapStart takes a snapshot of an already-initialized environment (bytecode loaded, class-loading finished, user init run) and clones it to serve many concurrent invocations.
Initial target was Java, where classpath scanning + JIT warmup dominated cold start; up to 90% cold-start reduction is the published target. (Source: sources/2024-11-15-allthingsdistributed-aws-lambda-prfaq-after-10-years.)
Architecture implications¶
- Per-clone isolated network. Each clone needs its own network
namespace with
tap,bridge,veth, and tunnel devices pre-provisioned before the VM resumes — the original Lambda on-demand network-creation path didn't fit. - Topology split at launch. On each host, the 2,500-micro-VM capacity was initially divided 200 snapshot-topology + 2,300 on-demand topology because snapshot networks require roughly 2× the Linux network devices per VM and per-device create/destroy cost grows with density. (Source: sources/2026-04-22-allthingsdistributed-invisible-engineering-behind-lambdas-network.)
- Forcing function for the 2022–2026 network retrofit. The decision to converge both topologies onto one worker and scale snapshot networks 200 → 4,000 per worker is what drove the Lambda networking team's constant-work, eBPF-NAT, and iptables-in-namespace fixes disclosed in the 2026-04-22 post. See patterns/pre-create-all-network-slots-at-boot.
Seen in¶
- sources/2024-11-15-allthingsdistributed-aws-lambda-prfaq-after-10-years — introduced as the Firecracker-snapshot unlock for Java cold-start, in the "tenets that held up" column of the PR/FAQ 10-year retrospective.
- sources/2026-04-22-allthingsdistributed-invisible-engineering-behind-lambdas-network — canonical disclosure that SnapStart's per-clone isolated-network requirement forced the Lambda network-topology retrofit: the original on-demand device-creation path didn't scale to snapshot's "pre-created and ready to attach" contract. The post discloses the 200-slot initial snapshot-topology cap and the sequenced 200 → 2,500 → 4,000 unification arc.
Related¶
- systems/aws-lambda — the product surface
- systems/firecracker — the snapshot substrate
- concepts/cold-start — the metric it optimizes
- patterns/pre-create-all-network-slots-at-boot — the network-side unlock that scaled the snapshot topology