Changelog
What shipped in each Cello release, from v0.6.3-ryuzcorp.4 to v0.2.0.
v0.2.0
Simple backward compatibility with
denoland/celld: swap the binary or
image on the same bucket. Stop every old node first — the peer protocol
is new (cello-tunnel, x-cello-overload, cello-log-stream, new
overload marker), so a mixed fleet is unsupported — then start 0.2.0.
Rename every CELLD_* variable to CELLO_*; the node refuses to start
otherwise, fail-closed.
Persisted identifiers are now written with the cello spelling and read
with both: __cello-runtime, the __cello$* stub envelopes (rows
written before the binding-isolation patches still decode as plain
data), cello-r2 metadata and cello-r2-cursor-* cursors,
cello-schema / cello-retention, and the cello-image: prefix. Data
an older node wrote carries over with no migration step and no rewrite
pass. In-process dunders and the wire identifiers were renamed outright.
See Compatibility.
v0.1.0
First Cello release. The renamed product: repo, crate, binary, and CLI
are cello; CELLD_* is CELLO_*; .celld/dev is .cello/dev; the
health path, telemetry names, and prose say Cello. Bucket objects, cell
SQLite, stored values, and the node-to-node wire stay celld-spelled
(see Compatibility with celld). Includes everything
below; no behavior change beyond the rename.
Releases v0.6.3-ryuzcorp.4 through .13 shipped as celld, so their
entries below use the historical spellings they shipped with —
CELLD_*, .celld/dev, and the celld binary — not the current
CELLO_* names.
v0.6.3-ryuzcorp.13
Adds 0079–0136. No bucket format bump and no wire protocol change
between nodes, so .12 and .13 nodes can share a bucket; the only new
bucket-side artifact is the log/<node>.loss.json loss record, which
older nodes read past, and a downgrade restores everything .13
wrote. Default behavior changes: the node-session id is pinned in the
data dir (0092), service.version carries the operator’s application
version instead of Cello’s build (0132), and WebSocket close codes
1001–1014 (other than the reserved no-status codes) now close instead
of throwing (0113).
New CLI: celld versions [SCRIPT] lists deployed versions (0123);
celld d1 --local runs execute, export, and migrations against the
running celld dev (0121); celld dev --state DIR (or
CELLD_DEV_STATE) keeps local state outside the project (0124).
- Runtime and cancellation:
0079,0080,0081,0083,0084,0085,0086,0087,0118settle shared, waiter-owned, and background work loudly instead of hanging;CELLD_MAX_CELLS_PER_ISOLATEcaps cells per isolate (default 32) (0135);CELLD_BACKGROUND_BUDGET_Sends post-response background work past the budget (default 30 s) without failing the answered request (0101); a poisoned cell restarts on a fresh isolate (0104);AbortSignal.timeoutfrees its slot (0106); the budget-orphaned workflow drive drops on retry (0125); failed covered alarms disarm while uncovered ones re-arm at scan pace (0127). - Fleet and durability:
0088–0096bound seal proof, renewal readbacks, alarm deadlines, tail streaming, lease conflicts, loss declaration, ensemble resumes, and wake timing;0103collects record-less fragments;0105prunes live epochs after unclean restarts;0126retries transient bucket write conflicts;0129routes dead-owner takeovers with capacity. - Storage:
0098,0105,0099,0122,0100cover compaction retention, unclean-restart pruning, the single-writer dev store, indexed local LISTs, and the repaired facet-image import;CELLD_SQL_STATEMENT_MSbounds per-statement time (unset means unbounded) (0128); hydration counting is O(1) (0102). - Workers compatibility:
0109–0117,0108,0107,0130,0113;0119refuses foreign-script_namebindings at deploy time;0131reserves the host-injectedCELLD_NODEname. - Telemetry:
0132–0134,0136. - New env knobs and their defaults:
CELLD_BACKGROUND_BUDGET_S(30),CELLD_MAX_CELLS_PER_ISOLATE(32),CELLD_SQL_STATEMENT_MS(unset),CELLD_TMP_MAX_BYTES(67108864, 64 MiB),CELLD_OTEL_LOGS(unset),OTEL_RESOURCE_ATTRIBUTES(unset),CELLD_DEV_STATE(PROJECT/.celld/dev),CELLD_NODE(generated and pinned tonode-idin the data dir).
v0.6.3-ryuzcorp.12
Adds 0076–0078. No API, CLI, or bucket format change, so .11 and
.12 nodes share a bucket and a downgrade restores everything .12
wrote.
- An eviction skips the full-database handoff snapshot when the cell’s
restorable chain was born in this epoch, every acked row is already
per-cell, and the chain holds at most 16 objects (
0078, modelled on ClickHouse MergeTree’s lazy merges). Measured on 200 agents × 6 writes, bucket posture, release builds: Class A 1934 → 1734 (−10%), PUT 1400 → 1200, write p50 unchanged at 210 ms. On 50 agents: Class A 582 → 532. Restored-from-older-epoch cells, group-commit rows not yet per-cell, and longer chains still snapshot as before. tools/s3benchdefaults to the release binary and refuses to start while another run is up (0076).- Test nodes boot on ephemeral ports, removing a port-reuse flake
(
0077).
v0.6.3-ryuzcorp.11
Adds 0058–0067, 0069–0072, and 0075 (0068, 0073, 0074
parked unmerged). No API or CLI change. A .11 node publishes bucket
format 3, which .10 nodes read, so .10 and .11 nodes share a
bucket.
- Bucket-posture group commit (
0059–0064,0070,0071) is in the binary but off by default (0075);CELLD_BUCKET_BUNDLES=1turns it on. It needs every live lease to read format 3 (every node on.11) and falls back to per-cell uploads at once otherwise. Measured on 200 agents × 6 writes: Class A 1934 → 1140 with grouping on, but write-ack p50 210 → 445 ms; on a hot cell, p50 235 → 432 ms. Enable it only where request cost matters more than write latency..10recovers a bucket.11wrote bundles into (whole-prefix fold, no loss declared), so a downgrade is safe. - An eviction proves a journaled epoch with one HEAD instead of a
prefix LIST (
0061). - The C heap is trimmed sooner after evictions (
0065,CELLD_C_TRIM_MB, default 8). Post-drain RSS fell ~10 MiB on the agent-50 shape. CELLD_SQLITE_CACHE_KBcaps the per-connection SQLite page cache (0066; default 2000 KiB is unchanged).- A deployment generation projects its
/bundletree once instead of once per isolate (0067). tools/s3bench(0058,0069,0072) measures bucket request counts and per-write latency against a local S3 for cost claims.
v0.6.3-ryuzcorp.10
Adds 0054–0057, all bucket-request and storage savings; no format
or API change, so .9 and .10 nodes share a bucket.
- A bucket-posture eviction makes fewer requests (
0054). An eviction that publishes a handoff snapshot then deletes, in the background, the per-cell L0/L1 objects the snapshot supersedes — one single-keyDeleteObjectper key (0057). The bucket credential needss3:DeleteObjectoncells/, which the read-write state grant already covers; a failed delete only leaves garbage behind and never fails the eviction. - Concurrent bucket-posture writes on one cell share ownership reads
(
0056): fewercells/<cell>/own.jsonGETs on hot cells at the same acknowledgement rule. - A fleet-posture node with no peer it could ever recruit backs its
nodes/scan off to the 30 s maintenance tick (0055), so a node that starts next to it is recruited within ~30 s instead of ~1 s (the new node recruits the old one itself at once).
v0.6.3-ryuzcorp.9
Adds 0052–0053.
- No behavior change with
CELLD_DEPLOY_AWS_*unset. A platform that sets the trio must give the deployment credential read access to bothdeploy/anddeploy-blobs/, and keep the ordinaryAWS_*grant read-write over the state prefixes (cells/,log/,nodes/,fleet/,wake/,drain/,kv/,r2/,probe/,telemetry/). The runtime also GETswake-format.jsonand LISTswake-v2/andwake-retired-v2/(one read-only prefix grant overwake-covers it), and lists and deletes thenode-cells/markers (that prefix stays read-write). The deploy command keeps the ordinary credential with write access to the deployment prefixes. A control-plane node, managed (CELLD_CLOUD) or BYO bucket alike, must not set the trio: it publishesdeploy/itself, so it refuses the trio at boot. - A serving process is non-dumpable: no core dump comes out of a node,
and a non-root debugger or profiler cannot attach (
ptracerefused); root tools are unaffected. Anything reading the node’s/proc/<pid>/environ,mem, ormapsas the fleet uid now getsEACCES(0053). The window betweenexecveandmainsetting the flag is still dumpable, and spawned children start with an emptied environment (PATH,HOME,TMPDIRonly), so no credential reaches them.
v0.6.3-ryuzcorp.8
Adds 0051: celld deploy --read-root DIR (repeatable) names
directories beyond the project that esbuild’s inputs and
containers[].image Dockerfiles with their build contexts may come
from. A platform that deploys a built subdirectory of a checkout, or a
monorepo app whose packages are hoisted above its config, names the
checkout. Without the flag, .7’s project-only bounds are unchanged.
v0.6.3-ryuzcorp.7
Adds 0043–0050.
celld deployandcelld d1refuse a project path that resolves outside the project directory — the config file’s own directory (0050). That covers the config,main, assets, Dockerfiles and build contexts,migrations_dir, Python sources, and every file esbuild bundles. A monorepo that hoistsnode_modulesabove that directory no longer bundles: install the packages inside it.migrations_direntries must be regular files, and acontainers[].imagethat is an absolute or..path is refused.- On Linux, esbuild runs under Landlock with reads confined to the
project, so
CELLD_ESBUILD(oresbuildonPATH) must be the self-contained release binary, not a wrapper script (0050). A kernel or seccomp profile without Landlock runs it unconfined with a warning; Docker’s, containerd’s, and podman’s default profiles allow it. celld r2 headandcelld r2 get --jsonprint a{"error":"not_found",…}object on stdout for an absent key (0044); exit code and stderr unchanged.celld --versionreports the full build version (0046).- New, opt-in:
CELLD_CACHE_SCOPE=deployment(0045);celld invokeruns one request against a staged version beforecelld activate(0048), with--header-envfor secret headers (0049);celld deploy --sourcemapcollects ano_bundleentry’s sibling map (0047). celld invokeneeds every node on this release. A stream received over RPC now stays open while the receiver holds it (0043), ending the spurious aborts of signal-carrying container execs.
v0.6.3-ryuzcorp.6
Adds 0009–0042.
- Upgrade every node of a fleet together. A rolling upgrade moves cells
with no migration (
0037), but cross-version traffic fails until every node runs this release: operator calls through an old node fail closed on new nodes, Queue DLQ and redrive forwards fail in both directions, and stubs sent from an old node revive as plain data on a new one (0040,0041). The Queue alarm replays the durable DLQ and redrive marks once the fleet converges. - Storage rows holding stubs written before the upgrade read back as
plain data (Durable Object stubs as null). Get each stub again from
its namespace or binding and store it again to write the new envelope
(
0040); storing the read-back value does not restore a stub. - Workers now fail when they reach a resource their script never
bound, touch another cell’s storage, call D1, KV, or Queue operator
methods through stubs, or use a socket id outside its owning cell or
request (
0022,0040,0041,0042). Declare every binding a Worker uses, and run operator actions through thecelloCLIs. - The Workers Cache API takes up to 64 MiB of memory per node by
default;
CELLD_CACHE_MAX_BYTES=0turns it off (0018). - Shared D1, KV, and Queue cells move to a runtime pool that
/statereports as__celld-runtime(0037). - Opt-in hardening stays off by default:
CELLD_EGRESSisopen(0021),CELLD_FENCE_NODE_NETNSisoff(0014),CELLD_MEMORY_SCOPEiscontainer(0011), andCELLD_OPERATOR_TOKENis unset (0013). - Set
CELLD_OPERATOR_TOKENonly once every node runs this release: older nodes send no token, so a mixed fleet with the token set fails internal calls with 401 (0013). Set it in the environment of everycelloCLI too; liveness and readiness probes use the public health endpoint. - Queue rows whose dead-letter source was forged before the upgrade can
still be redriven to the name they carry; planting such rows is
closed by
0041, redrive is operator-only, and retention ages them out.
v0.6.3-ryuzcorp.5
Adds 0007–0008 (durable Object RPC across isolates; container exec
exitCode survival).
v0.6.3-ryuzcorp.4
Patches 0001–0006: per-container image selection, exec waiting on
its start, node:crypto scrypt, streaming-body Response.clone(),
service-binding RPC stubs, and container exec pty. First release
published by the fork.