Changelog¶
User-facing changes to the runtime, newest first. Each day is grouped into Added, Changed, Fixed, Security and Performance. Internal refactoring and test-only changes are left out.
Version 0.1.0 is not released yet
The workspace version is 0.1.0 and there is no tagged release. Every entry below belongs to the unreleased 0.1.0 work on the development branch. When 0.1.0 is tagged, this page will gain a dated release heading.
0.1.0 (unreleased)¶
In progress¶
Added¶
- Operator SDKs as runtime services: the operator publishes an SDK (
POST /v1/sdks,21ci sdk publish), and tenants import its contract (import { acme } from "std:sdk",sdk:acme, WASMstd:host/sdk) while every call runs in the release their grants select (node, plan, tenant, project, environment; channels such asstableandcanary). An SDK update, rollback or removal never rebuilds or redeploys a tenant service.21ci sdk ls/types, MCPlist_sdks/get_sdk_contract. See Operator SDKs and bindings and Use operator SDKs. - Typed RPC for WASM components:
call-rpccalls a service's RPC method, and therpcexport (worldrpc-service) serves RPC calls, from Deno and WASM callers alike.21ci types <service>writes the.d.tsof a service's RPC surface. - Projects and environments:
/v1/projects,21ci project,--project/--envonservice createandpush; every tenant has adefaultproject. See Projects and environments. - Self-service tenant tokens: the
keys:writescope,POST|GET /v1/keys,DELETE /v1/keys/{key}and21ci token. A new token never holds more scopes, projects, environments or lifetime than its creator. See Get access. - An MCP server at
POST /mcpon the control listener, and the21ci mcpstdio bridge, so Claude Code, Cursor, Claude Desktop and other MCP clients can manage services with a tenant key. See AI assistants (MCP). scaling.min_instancesandscaling.max_instancesper service (--min-instances,--max-instances), bounded by the new plan keymax_instances; a service with a floor is kept warm from node start. The autoscaler also grows a service when its arrival rate rises fast, before requests queue. See Instances and autoscaling.std:sqlite(alsobun:sqlite): the service's SQLite database with bun:sqlite's synchronous API. See Synchronous SQL.- Atomic counters:
kv.incr(key, delta, { ttlSeconds? })instd:kvandincrinstd:host/kv. The TTL applies when the call creates the counter, so it doubles as a fixed-window rate limiter. See the guest API. - The client address of an HTTP request is available to handlers as
ctx.client.ipand in thex-rt-client-ipheader, resolved behindingress.trusted_proxies. - Control API:
GET /v1/me(who the credentials are, their scopes and the effective plan),GET /v1/version,GET /v1/plansandGET /v1/node/config(the running configuration, node secrets left out) for the operator, andGET /v1/services/{s}/hosts/{host}. A tenant key withlogs:readcan read its own tenant's usage. See the API reference. - Admission state per service:
status.admissionin the service object andadmissioninGET …/statsshow the adaptive concurrency limit, running and queued tasks, the queue wait and the autoscaler's view.GET /v1/node/scheduleradds the node-wide admission bounds and one view per service. See Scheduler and autoscaling. 21cicovers every control API operation: newwhoami,doctor,plan list|get,tenant update|usage,host attach|list|get|detach,data export,build-logs,versions <svc> N,service update --set|--unsetandnode config get; a global--retryfor busy answers,--jsonon every command, and exit code 4 for "busy, try again later". See the CLI reference.
Changed¶
- Operator bindings are reachable from operator SDK services only and act for the calling tenant (tenant header, plan,
rate limit). A tenant service calling
platform.fetchgets403 binding_not_allowed. -
The
[[sdk]]configuration section (SDK code bundled into tenant services) andpreloadwere removed; a configuration that still has them fails to load with a pointer to21ci sdk publish. -
Authorizationheaders now reach the handler unchanged, so a service can check its own tokens. Only a tenant API key (Authorization: Bearer rtk_…) is removed; clients ofaccess: keyservices that also need their own token send the invoke key inx-rt-key. status.inflightof a service reports the tasks running now (the same number asstatus.admission.running).- API keys are scoped by projects and environments instead of services:
project_idsandenvironmentsreplaceservice_idson/v1/keys,/v1/tenants/{t}/keys,/v1/meand21ci key/21ci token(--project,--environment). A key with a project or environment restriction cannot create, change or delete projects. Anenv:<name>call entry compares environments within the target's own project only. See Get access. - The metadata schema is one consolidated version. A
meta.dbfrom an earlier build is not migrated; see Upgrades for the reset. - The service-data (blocking) pool is work-conserving: when threads are idle, a busy tenant may use more than its
blocking_share, always leaving one thread free.
2026-10-07: review fixes and sample apps¶
Added¶
- Five sample apps in
sample/apps/of the runtime repository: a Telegram bot, event sign-ups, a live queue, UZS currency rates from cbu.uz, and a WASM edge gateway. Each one has a tutorial. std:host/platformfor WASM guests:fetch(binding, request)calls an operator binding with the same plan, method and path checks asplatform.fetchin Deno.- WASM components built with the full Rust standard library (which import
wasi:cli/terminal-*) now link and run. - Per-service Deno isolate reuse:
deno.reuse_isolatesin the service config, and21ci service create|update --reuse-isolates. - Limits for reused isolates:
deno.reuse_max_idle,deno.reuse_max_idle_total,deno.reuse_idle_ttlanddeno.reuse_max_uses.
Changed¶
- Requests may use ⅞ of
sched.max_inflight; the rest is reserved for service calls and builds. A tenant can hold at most half of the node's queue, and an explicitscaling.max_queuedis clamped to the service's share. Retry-Afteron503 queue_timeoutis now the estimated time for the queue to drain (1 to 60 s, with jitter). A client that disconnects while waiting gives up its place.scaling.max_concurrencyis now calledscaling.max_instances(the old name is still accepted);0means "no explicit cap".- The autoscaler runs on its own thread, and services with almost no demand get no warm instances.
- Changing a service's
environmentneeds the operator token or a key with no project or environment restriction, and is recorded as an event. - A service of another tenant can only be admitted in
call.allowed_callersby its service id. - Custom host claims keep their token and expiry when claimed again. After an unverified claim expires, the same
tenant must wait one more TTL (
409 host_cooldown), and a tenant can have at most 10 pending claims (hosts.max_pending_per_tenant). SIGHUPrefuses to drop a plan that tenants still use. While a control change is being applied, invokes may get503 route_stale(retry after 1 s).- A TCP internal listener now needs
listeners.allow_tcp_internal = true. - WASM guest clocks advance only after a real wait, in whole milliseconds.
Fixed¶
- Nested service calls no longer deadlock behind a tenant's concurrency cap, and eager builds no longer lift that cap.
- A killed or failed streamed response now aborts the body instead of ending as a clean
200. - Environment variables are fetched before the handler runs; a failure throws
env_unavailableinstead of blocking. - CPU time of WASM tasks with streamed responses is measured and billed again.
- Idle WASM instances share one node-wide slot budget, so one tenant's idle instances cannot exhaust the pool.
- A cancelled request rolls back its open SQL transaction; SQL result caps count the real in-memory size; a service database grows only within the tenant's remaining storage quota.
- Version cleanup keeps artifacts used in the last 30 minutes, and an hourly sweep removes leftovers of failed builds. A crash while rewriting the disk cache no longer looks like tampering.
Security¶
- Deno memory limits also cover resizable
ArrayBuffers, buffers the runtime creates (KV and fetch bodies,TextEncoderoutput) andBlobdata. Atomics.waitis disabled and shared WebAssembly memories are refused in Deno guests.- Guest SQL cannot create temporary tables, views or triggers.
- Failed-authentication limits group IPv6 clients by /64 and keep separate budgets for the control and public listeners.
rt secrets resealwrites with compare-and-swap, so a concurrent secret update is not overwritten.- Parked WASM instances refuse every
std:hostcall. Decrypted environment values are evicted after 10 minutes idle and when a tenant is disabled or deleted.
2026-10-06: queueing instead of 429, warm reuse, call access, faster request path¶
Added¶
- Requests over capacity wait for a slot instead of getting
429. Each service has a FIFO queue and tenants take turns; a request that waits longer thansched.queue_timeout(1 s inmin_latency, 5 s inbalanced, 10 s inmax_density) gets503 queue_timeoutwithRetry-After. The node bounds its queue withsched.max_queuedandsched.max_queued_mb(503 overloaded). See Scheduler and autoscaling. - Adaptive per-service concurrency, and warm pools that follow recent demand within
sched.warm_shareof the memory budget. - Opt-in warm isolate reuse for Deno (
deno.reuse_isolates, on by default inmin_latency). - WASM: per-service
wasm.reuse_instances, streamed responses,std:host/callandstd:host/queue, and a warm pool of pre-created instances. - Deno request bodies stream into the handler as a
ReadableStream. - Service environments (
environment, defaultproduction) andcall.allowed_callers; refused calls and enqueues get403 call_not_allowed. New CLI options21ci service create --env --allowed-callerand21ci service update. - Custom host verification by a DNS TXT record at
_rt-verify.<host>(hosts.verify_dns), and expiry of pending claims (hosts.pending_ttl). - Master key rotation with
RT_MASTER_KEY_PREVIOUSandrt secrets reseal. ingress.backlog(default 4096) and thert_ingress_conn_rejected_totalmetric.- A benchmark harness in
bench/.
Changed¶
- A plan's
allowed_triggersis enforced for HTTP (403 trigger_not_allowed), service calls, queue consumption and every schedule fire. WebAssemblyinside Deno is limited by real memory use: allocations overmemory_mbthrowRangeError, and wasm memories are sampled so a task over budget ends withoom.SIGHUPreloads[plans]instead of stopping the node.- Debug builds without
RT_ROOT_TOKENrefuse a control listener that is not on loopback. - Guest SQL results are capped at 10,000 rows or 10 MiB per call (
sql_result_too_large). - Uploads clean up old versions beyond
deploy.retain_versions, including vendored modules no version uses. - Disabling a tenant removes its services from routing and re-checks its cached keys.
blocking.threadsdefaults to the larger of 4 and the number of cores.- Plan
custom_host_suffixesmatch on a DNS label boundary. - The compiled-artifact disk cache moved to a new layout (
cache/v2/); older cache directories are removed at startup and services rebuild on demand.
Fixed¶
- Interval, cron and queue tasks keep their own default limits (300 s wall time, 10 s CPU) when the service sets none, instead of the HTTP limits.
- A late queue acknowledgement no longer settles a newer delivery attempt.
- Rebuilds use the version's
rt.lockand its vendored modules, so they reproduce the first build. Single-file JavaScript uploads are no longer rebuilt as TypeScript. - Async
WebAssembly.compileandWebAssembly.instantiatein Deno now complete. - A Deno task killed while waiting on I/O stops right away.
- Schedules keep their next due time across restarts.
Security¶
- The SSRF guard blocks 6to4 addresses by their embedded IPv4 address, Teredo,
192.0.0.0/24,198.18.0.0/15,240.0.0.0/4andfec0::/10, and adds the node's own interface addresses to its denylist. - Redirects from
httpstohttpare refused unlessegress.allow_https_downgradeis set. Outbound connections are pooled per tenant and per validated address, never shared between tenants. - One failed-authentication limiter covers the control API and
access: "key"invokes, keyed on the real client address behind trusted proxies. - The internal Unix socket is created with mode 0600 and checks the peer's uid (
listeners.internal_allowed_uids).
Performance¶
- No metadata database round trip per invocation: env, decrypted secrets, plans and egress rules are cached in memory, and the node needs far fewer threads under load.
- Concurrent readers for the metadata and service databases, and SQLite group commit for KV and blob index writes.
- The Deno isolate pool is refilled between tasks, env is read on first use, and the V8 code cache is stored in a binary format.
- The build fetch cache is capped at 64 MiB and
build.max_deps_mblimits a build's total dependencies. Cache access frequencies survive restarts, and the hottest services are preloaded at startup. Logs are written in batches.
2026-10-02: first implementation¶
Added¶
- The
rtbinary: the/v1control API, the invoke router, OpenAPI 3.1 with a Scalar UI, drain, andrt openapi. - A Go-style scheduler with elastic processor kinds, work stealing, a monitor thread and a deterministic simulator.
- The Deno executor: a fresh isolate per task from a build-time snapshot, the
std:*modules, a frozen clock and CPU, memory and wall-time limits. - The WASM executor: wasmtime components with a pooling allocator, epoch-based preemption and
std:host/{log,kv,sql,blob}. - The build pipeline: archive-safe validation, import resolution, guarded dependency fetches with
rt.lock, deterministic bundles and WASM precompilation. - A tiered cache (memory, disk, source) with cost-aware eviction and integrity-checked disk entries.
- Storage: the metadata database, a content-addressed artifact store, per-service KV, SQL and blob storage, and a durable queue with a dead-letter queue.
- Triggers: the route table, interval and cron schedules, the queue engine, service calls and rate limits.
- Security: tenant API keys with scopes, AES-GCM sealed secrets, quotas, the SSRF guard, and the memory governor.
- Observability: the log store with SSE streaming, latency statistics, Prometheus metrics and scheduler traces.
- The
21ciCLI for contexts, tenants, keys, services, push, rollback, env and secrets, schedules, queue, logs, invoke and node operations. - Dockerfile, Compose file, scripts and Makefile; runnable samples; user guides for the API, handlers and deployment.
Fixed¶
- Per-client bursts in ingress rate limits are honored.
- The
rt.jsonimport map is applied during builds. - A queue retry delay is honored instead of the message being acknowledged.
- Cross-tenant calls work when the plan allows them.
- Waiting requests resume only after the build result is recorded; called services receive the base-path and trigger headers.
- The log filter reads
RT_LOGfirst, thenRUST_LOG. - The configuration tools match the runtime's loader for every profile.
2026-10-01: specification and configuration tools¶
Added¶
- The design specification, a security and performance review, and the implementation plan.
- The configuration schema, profile defconfigs (
max_density,balanced,min_latency) and themake menuconfigterminal UI.