Changelog

New features, improvements, and fixes shipped to E2B.

​

May 11, 2026

sandboxsdk

New features

• Snapshot compression. Memory and root filesystem snapshots are now compressed with zstd, reducing storage size and pause/resume bandwidth. Older snapshots stay compatible — the read path auto-detects the format. Rolling out gradually behind per-team/cluster/template flags.
• Custom domain header transforms. Sandboxes can now be configured with header injection rules at creation time, and the configuration persists across pause and resume. See the custom domain docs.
• Stable per-sandbox hostname routing. SDKs can now reach sandboxes through a stable sandbox.<domain> host using header-based routing instead of per-sandbox hostnames, simplifying egress, DNS, and TLS setup.
• HTTP/2 in the Python SDK. The Python SDK now talks to the E2B control plane and envd over HTTP/2, reducing connection overhead under concurrent calls.
• HTTP/2 in the JavaScript SDK. The JavaScript SDK now uses HTTP/2 for sandbox traffic on Node ≥ 20.18.1. Non-Node JS runtimes are unaffected.

Updates

• Faster pause/resume on snapshot storage. Redundant snapshot-header fetches dropped from thousands to a couple per operation, lowering tail latency and storage cost during pause, resume, and template builds.
• Lower latency on connect under contention. The sandbox connect endpoint now uses pub/sub wake-ups with jittered backoff instead of constant polling, reducing tail latency when many requests target the same sandbox.
• Smaller snapshot diffs. Writes that zero or discard disk blocks are now omitted from snapshot diffs and reclaimed from the underlying cache, shrinking snapshots and freeing host space sooner. Existing snapshots remain compatible.
• Smaller diffs for small-file workloads. New template builds enable ext4 inline data, so files under ~160 bytes live inside their inode. Shrinks snapshot diffs for workloads that churn many small files. Applies to newly built templates.
• Lower snapshot growth from guest memory compaction. Newly built base templates disable proactive in-guest memory compaction, avoiding spurious dirty pages in the next snapshot. Existing templates pick this up on rebuild.
• Opt-in pre-pause memory and disk reclaim. Before pausing, sandboxes can optionally run sync, drop caches, compact memory, and fstrim to shrink the resulting snapshot. Disabled by default; rolling out gradually with per-step time limits.
• Free-page reporting on supported VMs. Sandboxes on newer Firecracker builds can release freed guest memory back to the host (rolling out behind a flag), reducing host memory pressure and snapshot size.
• Controllable kernel version for new template builds. The kernel used for new template builds can now be rolled forward at the platform level without a code change. Existing templates and snapshots keep their pinned kernel.

Bug fixes

• SDK connectivity over IP hosts. Fixed a regression where the proxy returned HTTP 400 when the SDK connected via a raw IP host (for example on-prem or e2b-local). Header-based routing now works correctly for IP hosts again.
• Dropped command output on disconnect. Fixed a deadlock in the in-sandbox process service that could freeze command output for all subscribers and hang new connect calls when one client went away. Shipped in envd 0.5.16.
• Crash on client disconnect during streaming. Fixed a nil-pointer crash that could occur when a client disconnected mid-stream from a streaming command.
• Reads after resume on compressed snapshots. Fixed an issue where post-resume reads could resolve to a non-existent uncompressed path with the new compression rollout. The base path is now resolved per call from live snapshot metadata.
• Leaked sandbox VMs. Fixed a case where stale VM processes could survive sandbox shutdown after rapid checkpoint bursts. The orchestrator now force-kills the cgroup as a backstop.
• Snapshot-upload race across orchestrators. Fixed a race that could surface stale snapshot data across hosts during pause/resume. Peers now always read the canonical finalized header from storage.
• NBD connection drop on trim/sync writes. Fixed a parsing bug that killed the NBD connection any time the guest kernel set request flags (for example FUA on a sync write). Production was unaffected because those flags weren’t advertised, but enabling them is now safe.