stalwart-migrator/ARCHITECTURE.md

stalwart-migrator — Architecture

Status: design, no implementation yet. Scope: upgrade a Stalwart Mail Server in place from 0.15.5 to the current latest release (0.16.14 as of 2026-08-19) with no data loss, a working a recovery point the operator provides, and an automated post-migration validation pass.

1. Why this isn't a thin wrapper

Stalwart does not ship an automated upgrade tool today (planned for 1.0, targeted H1 2026, not yet released). 0.15.5 → 0.16.x is a major boundary, not a patch bump, and it is unusually dangerous to automate naively:

• The v0.15 → v0.16 config model changes completely: multiple TOML files plus DB-resident settings collapse into one config.json that describes only the datastore connection, with everything else moved into JMAP-managed objects.
• Account names change from bare usernames to full email addresses; DAV URLs change (/dav/cal/alice → /dav/cal/alice%40example.com).
• On first v0.16 start, the server irreversibly deletes all directory records (users/groups/domains/tenants/OAuth clients), all settings, DMARC/ TLS/ARF reports, pending tasks, telemetry, spam training samples, and quota counters. Mail/calendar/contact data is untouched, but everything else is gone unless captured first.
• Migration requires a manual "recovery mode" boot of the new binary, then an external tool (stalwart-cli apply) replays a converted settings snapshot into it over HTTP while it's up in that special mode — a multi-process, multi-terminal, stateful procedure with no built-in resumability.
• In a cluster, every node must be stopped before migration starts; one node left on v0.15 corrupts the shared store.
• Real-world failure mode already reported in the wild: post-migration WebUI login breaks because the UI now requires HTTPS via defaultHostname, not plain IP access — a config/DNS issue, not a data issue, but it reads as "the migration broke everything" to an operator.
• Not all settings migrate automatically: SMTP listeners, routing, rate limits, spam rules, and auth backends are explicitly not carried over by Stalwart's own conversion script and must be recreated or replayed from a separately captured snapshot.

None of this is exotic — it's exactly what Stalwart's own UPGRADING/v0_16.md and resources/scripts/migrate_v016.py already do. This project's job is to turn that fragile, manual, two-terminal runbook into a single supervised, checkpointed, reversible operation — and to keep working as new releases land on top of 0.16.x, most of which (0.16.1–0.16.14, per changelog) are pure patch/feature releases with no schema migration, i.e. a binary swap + smoke test, not a full migration.

2. Design goals / non-goals

Goals

• Zero data loss for mail, calendar, and contact content (the one thing Stalwart itself guarantees is untouched — everything else is on us).
• Nothing destructive happens until the operator has confirmed a recovery point exists. This tool does not implement the undo (see the non-goals and §4.8); it refuses to start without being told one is in place.
• Fully automated happy path; the operator answers a preflight confirmation once, then watches (or walks away and checks the report).
• Resumable: if the process dies mid-migration (crash, SSH drop, OOM), a re-run picks up from the last completed checkpoint instead of redoing or, worse, double-applying destructive steps.
• Works across the deployment shapes Stalwart actually supports: systemd + bare binary, Docker/Compose, and single-node vs. cluster — with embedded (RocksDB/SQLite) or external (PostgreSQL/MySQL/FoundationDB) stores.
• Extensible to future major boundaries (0.16 → 1.0 and beyond) without a rewrite: version-boundary logic is pluggable, not hardcoded into the core engine.

Non-goals

• Not a recovery tool. Restoring a failed migration is the operator's own snapshot or backup, by whatever method they already trust — ZFS/LVM/ btrfs snapshots, VM or volume snapshots, or a restorable backup. This tool does not take one, verify one, or restore from one. §4.8 explains why that turned out to be the right split.
• Not a general Stalwart config management tool (no drift detection, no day-2 ops beyond the migration window).
• Not a replacement for routine backups — it produces a migration-time backup as a side effect, but ongoing backup policy is the operator's job (Stalwart's own guidance: import/export is explicitly not a backup substitute; Vandelay per-account export is the documented backup tool).
• Not a cross-major-version skip tool. If the source is older than 0.15.x, the tool requires stepping to 0.15.x first (this matches Stalwart's own stated constraint — see UPGRADING notes).
• No support for editing mail content during migration (no format conversion beyond what Stalwart's own store migration does).

3. High-level flow

 ┌─────────────┐   ┌───────────┐   ┌────────────┐   ┌───────────────┐   ┌────────────┐   ┌────────────┐
 │  PREFLIGHT  │──▶│  BACKUP   │──▶│ STAGE NEW  │──▶│ RECOVERY-MODE │──▶│  CUTOVER   │──▶│  VALIDATE  │
 │  (checks,   │   │ (defense  │   │  BINARY +  │   │   MIGRATE     │   │ (swap, up, │   │ (functional│
 │  rehearse)  │   │ in depth) │   │  config    │   │ (apply plan)  │   │  smoke)    │   │  + counts) │
 └─────────────┘   └───────────┘   └────────────┘   └───────────────┘   └────────────┘   └────────────┘
        │                 │                │                 │                 │                │
        └─────────────────┴────────────────┴─── on failure ──┴─────────────────┴──▶  STOP + REPORT
                                                                                     (operator restores
                                                                                      their own snapshot)

Each box is a phase; each phase is a sequence of idempotent, checkpointed steps. State is persisted to disk after every step (§5), so the whole pipeline can be killed and re-invoked safely.

For a pure patch bump within 0.16.x (no schema change per Stalwart's changelog through 0.16.14), the plan collapses to: PREFLIGHT → BACKUP → STAGE → CUTOVER → VALIDATE, skipping the recovery-mode phase entirely (see §4.6).

4. Phases

4.1 Preflight

Read-only. Aborts before touching anything if a hard blocker is found; warns and asks for confirmation on soft blockers.

• Detect current Stalwart version (stalwart --version, or JMAP Core/echo/session endpoint if remote).
• Refuse to run if current version is outside the tool's supported starting range (must be ≥0.15.0; older installs are told to upgrade to 0.15.x first, per Stalwart's own guidance).
• Detect topology: systemd unit vs. Docker container vs. Compose, single node vs. cluster member count (via config/cluster settings), reachable peer nodes.
• Cluster gate: refuse to proceed unless every node in the cluster is confirmed stopped (mirrors the documented hard requirement — one live v0.15 node during migration corrupts the store).
• Detect store backend(s): RocksDB, SQLite, FoundationDB, PostgreSQL, MySQL, plus configured blob store (local FS / S3-compatible) and FTS backend (native / Elasticsearch).
• Disk space check: require free space ≥ N× current data directory size (embedded stores need a full copy for the backup step; default threshold configurable, hard-fail below a safety floor).
• Resolve and download the target binary/image, verify checksum/signature against the published release.
• Fetch and pin the exact upstream migrate_v016.py (or its 0.16-successor equivalent) revision, hash it, vendor the hash into the run's checkpoint record — we depend on it as an external, versioned dependency, not a static local copy that can silently drift from upstream.
• Dry-run the settings dump against the live server (read-only JMAP calls) to confirm admin credentials and API reachability before anything destructive is scheduled.
• Snapshot pre-migration facts used later for validation: account count, per-account mailbox message counts (IMAP STATUS), domain list, DKIM key fingerprints, TLS cert fingerprints, listener port list. Stored alongside the checkpoint, not derived after the fact.
• Emit a plain-language plan summary and require explicit confirmation (--yes to skip interactively, but never by default).

4.2 Backup — defense in depth

No single backup mechanism is trusted alone, because the risk profile is different at each layer:

1. Filesystem/DB snapshot (infra-level, fast, whole-store):
   • Embedded (RocksDB/SQLite): stop-the-world cp -a of the data directory (or LVM/ZFS snapshot if available — preferred, since it doesn't require the copy to finish before the next step) to a sibling path (<datadir>.v0155-backup), never overwriting source.
   • External SQL (Postgres/MySQL): targeted dump of the critical table set Stalwart's own guide calls out (s d r h b g j f u for Postgres; equivalent for MySQL), not a full-instance dump — matches the documented, tested restore path and stays fast on large installs.
   • FoundationDB: fdbbackup against the configured cluster.
2. Settings/principals export (the migrate_v016.py dump step): captured during preflight and re-captured immediately before cutover, so the export used for the apply reflects the last-known-good state, not a stale preflight snapshot if time has passed.
3. Per-account content export (Vandelay/JMAP): for installations under an operator-configurable account-count threshold, take a belt-and-suspenders full vandelay import (i.e. export-to-file) of every account into self-contained per-account SQLite archives. This is independent of storage backend and of the in-place migration path entirely — if everything else somehow goes wrong, mail content is recoverable via Stalwart's own documented import path into a clean instance. Skipped above the threshold by default (time cost), but available as --full-content-backup regardless of size.
4. Binary preservation: old binary is moved aside (stalwart.v0155), never deleted, so putting the machine back by hand doesn't depend on re-downloading a specific old release under pressure.

Every backup artifact is checksummed and the checksum recorded in the checkpoint file. Before moving past this phase, the tool verifies the filesystem backup by opening it read-only with the old binary in a throwaway temp directory and confirming it reports the expected version and a sane account count — catching a corrupt or partial copy while the pre-migration instance is still up, rather than after it isn't.

4.3 Stage

• Install target binary alongside the old one (never overwrite in place).

• Run migrate_v016.py convert against the fresh dump to produce config.json + export.json, applying path rewrites for Docker/volume layouts detected in preflight.

• Additionally generate an apply-plan for the settings Stalwart's script does not carry over — SMTP listeners, routing rules, rate limits, spam rules, auth backend config — by diffing the old effective config against the new schema and emitting a best-effort JMAP object set for stalwart-cli apply. This is flagged clearly as best-effort and included in the final report for manual review; silently getting it wrong (rather than flagging it) would be worse than not attempting it.

  This is no longer an optional enhancement. Measured against a real production instance, migrate_v016.py migrated 219 of 12,401 settings — 1.8% — leaving 12,182 for the operator to recreate by hand, including server.listener. A migrated instance therefore serves nothing until somebody rebuilds its listeners, whatever else went right.

  Status: started (internal/applyplan), generating NetworkListener objects from server.listener.* and reporting its own coverage. Listeners came first because every other unmigrated setting degrades the server while this one stops it being a server at all.

  Two rules the package holds to:

  • Only mappings confirmed against a real v0.16 binary go in. The published schema reference gives NetworkListener.bind as a JSON array; 0.16.14 rejects that. The encoding it accepts — a value-keyed set, {"[::]:25": true} — was found by applying a plan to a live recovery-mode instance and reading it back with stalwart-cli snapshot. An unverified guess here produces a plan that fails at apply time or, worse, quietly configures the wrong thing.
  • Coverage is reported, never implied. Against the smoke instance the generator covers 24 of 3,505 unmigrated keys and says "0.7%", listing the largest groups it did not touch. A plan that covered a fraction while implying completeness would be worse than no plan.

  Operations are emitted as upsert with matchOn: ["name"], so a plan can be re-run — an operator will run it more than once — and the supplement is applied after export.json rather than merged into it, so a generated mapping can never override one the official script got right.

• Stage new systemd unit / Compose file changes without activating them.

4.4 Recovery-mode migration

This is the phase most exposed to partial-failure — it drives an external process (the new Stalwart binary) through an undocumented-duration startup, then drives a second external process (stalwart-cli apply) against it over HTTP. Both are supervised with explicit timeouts and health polling, not fire-and-forget:

1. Stop the old service.
2. Start the new binary in the foreground with STALWART_RECOVERY_MODE=1 and a freshly generated one-time STALWART_RECOVERY_ADMIN credential (random, never the operator's real password, never logged).
3. Poll the recovery HTTP endpoint until healthy or a timeout elapses; on timeout, capture logs and stop rather than hanging indefinitely.
4. Run stalwart-cli apply --file export.json, then the generated best-effort settings plan from §4.3, capturing full output.
5. Verify the apply reported success for every object (the tool parses the apply-tool's structured output rather than trusting exit code alone — partial application with a zero exit code is exactly the kind of silent failure this tool exists to catch).
6. Stop recovery mode cleanly (SIGTERM, not SIGKILL, to let it flush).

Checkpointed after each numbered step, so a crash between "apply succeeded" and "recovery mode stopped" resumes at step 6 instead of re-running apply against an already-migrated store.

4.5 Cutover

• Update the real systemd unit / Compose config to point at the new binary and config, without the recovery env vars (leaving STALWART_RECOVERY_MODE=1 set is a documented footgun — it would recovery- boot on every restart).
• Start the service normally.
• Wait for healthy JMAP session response.
• Trigger disk-quota (and tenant-quota, if multi-tenant) recalculation via the management API, and poll the task queue until it completes rather than firing and moving on.

Status: implemented (internal/cutover), but nothing calls it yet — see §8. Notes on how it turned out:

• It refuses to run at all unless the operator has confirmed a recovery point exists (§4.8). That's an acknowledgement, not a check — this tool can't verify someone else's snapshot — but it makes the irreversibility of this phase impossible to walk into unasked.
• The unit is rewritten in place, not generated from a template: an operator's unit carries hardening options, limits and dependencies this tool has no business having an opinion about, and regenerating it would silently drop them. It repoints ExecStart (preserving systemd's -@:+! prefix characters and every argument after the executable), updates --config if asked, and strips recovery-mode Environment= lines. It refuses on a unit with no ExecStart, and on an Environment= line that mixes a recovery variable with others — a line it only partly understands is one it must not edit.
• The original unit is preserved and recorded as the service-unit artifact before the rewrite, so an operator restoring by hand isn't reconstructing a unit file from memory.
• Docker deployments recreate rather than rewrite, because a container cannot be edited in place the way a unit file can. That makes silent loss the default failure: a container rebuilt without its capabilities, its custom network or its device mappings starts cleanly and is quietly not the server it was. So the same rule the unit rewrite follows applies here - a definition this only partly understands is one it must not rebuild - and cutover refuses a container using anything outside the set it carries across, naming what it found. The list is conservative and deliberately not exhaustive; docker's HostConfig has far more fields than it checks, and one it does not know about is a reason not to be recreating that container at all.
• The old container is renamed rather than removed, and the old image is never pruned. Together they are the container's manual restore path, the nearest equivalent to the preserved binary of §4.2: one command starts the previous container again. The docker inspect of the container as it was is preserved as the container-definition artifact before anything is replaced, for the same reason the unit file is.
• Status: the container path is implemented and not yet reachable from the CLI - run does not pass container options, so a container is still refused there. Wiring it up, and lifting preflight's refusal for the containers it can now handle, is the remaining work in #3.
• Quota recalculation is the one step allowed to fail without failing the phase. Stale counters are an accounting problem; a failed cutover is one an operator has to respond to by restoring a machine that is otherwise migrated and serving mail correctly. Calling for that over a counter would be the worse outcome, so it warns and points at the WebUI's Tasks panel.

The quota call itself is grounded in Stalwart's x:Task schema reference (docs/ref/object/task/), not guessed: x:Task/set creating one AccountMaintenance variant per account with maintenanceType: "recalculateQuota", exactly as the WebUI's own "Recalculate disk quotas" fans out. The upgrade guide only documents the WebUI path, so two details remain unconfirmed against a live server and are called out in internal/stalwartapi/task.go: whether the schema's "read-only" annotation on accountId/maintenanceType means "immutable after creation" (it has to, or the variant couldn't be created), and whether a finished task simply leaves the queue (TaskStatus documents Pending/Retry/Failed with no success state). That uncertainty is the reason this step warns rather than fails.

4.6 Patch-bump fast path

For an already-0.16.x install moving to a newer 0.16.x patch (the common case after the initial major migration, and per the changelog the case for every release from 0.16.1 through 0.16.14 so far): preflight confirms no schema-migration flag is set for the target version, and the plan skips §4.4 entirely — binary swap, restart, same validation suite as §4.7. This is intentionally the same engine with a shorter plan, not a separate code path, so it doesn't rot independently.

4.6a What the converter drops without saying so

migrate_v016.py consumes every acme.* setting and emits nothing for it, and does not list those keys as unmigrated either. The effect is a migration that reports complete success while quietly removing certificate renewal: the certificate itself carries over, so nothing looks wrong until it expires about ninety days later. Observed on a production migration, 2026-08-25.

The supplemental plan already generates what the converter leaves behind for listeners (§4.6). An AcmeProvider generator belongs alongside it. The object shape is known-good, having been applied to a live 0.16.19 server: challengeType and renewBefore are enums (TlsAlpn01, R23), contact is a value-keyed set rather than a list, and accountKey/accountUri are server-set and must be omitted — the server registers a fresh ACME account, since the v0.15 account key cannot be carried across.

4.7 Post-migration validation

What this suite can assert depends on the boundary being crossed, and on the 0.15/0.16 boundary it is less than this section originally claimed. Stalwart 0.15.x reports no per-mailbox message counts at any endpoint, and the impersonation login 0.16 offers returns 401 there, so there are no "before" counts to compare against — the before/after message-count comparison is simply unavailable for the migration this tool exists to perform. Both versions report per-account used quota, which is captured on both sides, but §4.5 notes the migration resets quotas to zero pending recalculation, so it is recorded rather than asserted on. What remains checkable across the boundary is that every account and every domain survived, and the reports say so in those words rather than implying a no-data-loss guarantee that was not measured. See internal/validate/content_integrity.go.

Runs automatically after cutover, against the service cutover has just started: that is the instance people will actually use — its real config, its real ports, under its real service manager — and checking it costs no extra downtime, where booting a second copy inside the maintenance window would. Failure stops the run, reports loudly, and exits non-zero, leaving the operator to decide what to restore (§4.8). The service is deliberately left running: by this point the store has been migrated in place, so stopping it undoes nothing.

A check that could not be performed is reported as skipped, never as a pass. Preflight only captures the "before" snapshot when it has an admin URL to capture it from, and a run without one has to say it compared nothing rather than imply everything survived.

internal/validate.BootCheck remains the equivalent for an instance the tool boots itself, which is what rehearse needs; run uses RunLive.

Of the checks listed below, what exists today is the account/domain comparison. The rest are the intended shape of the suite, not a description of it.

• Version check: reported server version matches the target exactly.
• Auth check: WebUI login succeeds over the configured hostname via HTTPS (not bare IP) — this directly targets the real-world post-0.16 login failure mode found in the field.
• Protocol reachability: JMAP session, IMAP, SMTP (submission + MTA), POP3, ManageSieve, CalDAV/CardDAV endpoints all accept a handshake on their configured ports.
• Directory integrity: account/domain/group counts match the preflight snapshot exactly (accounting for the bare-username → email-address rewrite, which the tool resolves by comparing normalized identities, not raw strings).
• Content integrity — the core no-data-loss check: per-account IMAP STATUS (MESSAGES) compared against the preflight snapshot for every mailbox of every account (or a statistically sampled subset above a configurable account-count threshold, with the full sweep always available via --full-validation). Any mismatch is a hard failure.
• DKIM/TLS check: key fingerprints and cert validity match or are intentionally rotated (new-key generation is an expected v0.16 behavior, not a bug — the check distinguishes "changed as documented" from "missing").
• DNS check: for domains under Stalwart's automatic DNS management (new in 0.16), diff expected vs. actual published records and flag drift rather than assume the automation ran correctly.
• Mail-flow smoke test: send one real message through SMTP submission to a dedicated canary mailbox and confirm it's retrievable via IMAP within a timeout — the one end-to-end check that nothing upstream can fake.
• Quota check: recalculation task (§4.5) completed and reported numbers are non-zero/sane where preflight showed non-zero usage.

Output is a single structured report (JSON + human summary): pass/fail per check, with enough detail to hand to the operator deciding whether to restore.

4.8 Recovery from a failed migration — out of scope

This tool does not undo a migration. Recovery is the operator's own snapshot or backup, taken by whatever method they already trust and know how to restore: ZFS/LVM/btrfs snapshots, a VM or volume snapshot, or a restorable backup. This tool does not take one, verify one, or restore from one. Cutover refuses to start until the operator confirms one exists (§4.5).

This replaced a working, tested rollback implementation, and the reasoning is worth recording because the deleted code looked good:

• Restoring bytes correctly is not the hard part; restoring everything else is. The implementation copied file contents and permissions and verified every restored file against a manifest — and did not preserve ownership. Run as root, it produced a byte-perfect, checksum-verified, root-owned data directory that Stalwart, running as its own user, could not open. It would have reported success. A filesystem snapshot has no such failure mode, because it never lost the metadata in the first place.
• The external-database path was worse. pg_dump without --clean emits CREATE TABLE + COPY; replaying that into a database whose tables still exist fails outright, and ON_ERROR_STOP=1 — added so a half-applied restore couldn't be reported as success — turned that into a hard failure. The two SQL paths were asymmetric and only one was plausibly correct.
• It was never exercised against anything real. Every test drove fake systemctl, psql and stalwart binaries. That's sound for logic and ordering, and it is not evidence about production.
• Snapshots are already in the operator's runbook. They are atomic, metadata-preserving, cheap with copy-on-write, and cover the whole system — binary, unit file, config, data — rather than the subset one tool thought to capture.

What this tool keeps doing, so a manual restore is as easy as possible:

• The old binary is preserved next to the new one (§4.2), never deleted.
• The original service definition is preserved as a service-unit artifact before cutover rewrites it, so the operator doesn't reconstruct a unit file from memory.
• Every artifact path and checksum stays in the checkpoint, and status <run-id> prints exactly which steps completed and which failed.

The mail-delivery gap is accepted. Restoring any pre-migration recovery point discards mail delivered since it was taken. This was equally true of the rollback implementation, is inherent to restoring a point in time, and is not something this tool can solve. Plan the migration window accordingly.

Two consequences worth being explicit about. First, the confirmation cutover requires is an assertion, not a check — an unverifiable promise is weaker than a guarantee, and the value is only that nobody migrates a production mail server having never been asked the question. Second, there is no longer an automatic response to a failed migration: a failure stops the run and reports, and a human decides what to restore. Both are deliberate trades for not shipping a recovery path that has never been tested against a real server.

4.9 Rehearsal (was: dry run)

This section was rewritten after running the previous design against a real 0.15.5 instance and a real production settings corpus. What it found inverted the design's assumptions, so the reasoning is recorded here rather than quietly replaced.

The original dry run existed to answer "will the migration mechanics work?" — it cloned the data, ran the real recovery-mode migration against the clone, booted the result, and compared content before and after. Three findings retire that design:

1. The mechanics were never the risk. Backup, settings dump, convert and the recovery-mode store migration all worked essentially first time against real software. The failures were everywhere else.
2. The final comparison cannot work, at all. It needs the migrated sandbox to answer an API. server.listener is not among the settings migrate_v016.py migrates, so a migrated instance has no listeners and answers on nothing. That is not a sandbox artifact to engineer around — it is the true post-migration state.
3. The expensive half buys the least. Against a 3.6 GB production store the old flow copies the data twice (backup + sandbox clone, ~11 GB and a long wait) while reading a live mail store, to prove that RocksDB files copy correctly and that recovery mode can open them. Real, but modest.

Meanwhile the cheap half — dump, convert, and report what did not convert — is what caught every problem that would have derailed a real migration: an empty defaultHostname that v0.16 rejects, accounts whose passwords v0.16 refuses to create, and a reconstruction worklist of 12,182 settings. It needs no data copy at all.

So the phase reduces to the half that earns its cost:

stalwart-migrate rehearse. Run preflight (§4.1, read-only), dump settings and principals from the live instance, run migrate_v016.py convert, and report:

• the generated export.json plan (what will carry over), and
• unmigrated.txt (what will not, grouped and counted — see §4.3).

That is the whole phase. It copies no data, clones nothing, starts no server, and never writes to the store — so it is safe to run against production repeatedly, early and often, without a maintenance window. It answers the question that actually decides a migration plan: what will I have to rebuild by hand, and does my configuration convert at all?

The sandbox is gone. Cloning the store to run a migration against the copy proved only that the store migrates and opens — which is worth something, but not the disk and the wait, and not the risk of reading a live store to get it. Where that assurance is wanted, rehearse the whole thing on a throwaway VM restored from a backup, which is what the smoke environment already does and does better.

Consequences worth stating, since they make this phase much cheaper than its predecessor:

• No target binary is needed. convert is pure Python; nothing in this phase executes a Stalwart binary of either version.
• No disk headroom is needed. Nothing is copied. Preflight's free-space check still runs, but it is anticipating the backup a real run will take, not anything rehearse does — and it says so.
• Rehearsal performs no content-integrity comparison. §4.7 explains why that is unavailable on this boundary regardless of how it is staged.

Artifacts live under work-dir/<run-id> and are removed on every exit path unless --keep-artifacts is passed — with one deliberate exception. unmigrated.txt is the operator's reconstruction worklist and is preserved and checksummed even on a clean run, because deleting it would throw away the most useful output of the whole exercise.

A same-boundary patch bump (§4.6) needs no settings conversion at all, so rehearse for that plan reports that there is nothing to rehearse.

5. State machine / checkpointing

Every run gets a run-id and a checkpoint file (/var/lib/stalwart-migrator/runs/<run-id>/state.json) written after each step completes, containing: run-id, source/target version, current phase/step, timestamps, artifact paths + checksums, and the preflight snapshot facts used by validation. Steps are pure functions of (checkpoint-state → new-state); re-invoking stalwart-migrate run with an in-progress run-id resumes at the first incomplete step. Steps are written to be safe to re-run if they were interrupted mid-execution (e.g. the filesystem copy step checks for and resumes/redoes a partial copy rather than trusting a checkpoint that says "started" as if it meant "done").

This is the same shape as a deployment pipeline's state file, deliberately — the risk profile (long-running, multi-process, must survive being killed) is the same problem.

6. CLI surface

stalwart-migrate preflight   [--config PATH] ...              # read-only, prints the report
stalwart-migrate rehearse    [--keep-artifacts] ...            # read-only; see §4.9
stalwart-migrate run         (refused today — see §8)
stalwart-migrate status      [run-id]                          # implemented
stalwart-migrate report      <run-id>   [--json]                # not yet implemented

run is the only command that mutates anything: rehearse reads the live instance and writes only inside its own work directory. run always starts with preflight. Nothing in this tool restores a failed migration (§4.8), so there is no rollback command, and no confirm step to close a rollback window that no longer exists.

The migration-time artifacts a run leaves behind — the preserved old binary, the settings and principals dumps, the preserved service definition, and and rehearsal's converted plan and unmigrated worklist — are never pruned automatically. They're small next to the data directory, they're what a manual restore reaches for first, and deleting them on a schedule to reclaim disk would be the tool making a call that isn't its to make. Flags shown here are the design intent; run stalwart-migrate <command> -h for the actual current flag set.

7. Project layout (Go, matches this workspace's other CLI tools)

stalwart-migrator/
  cmd/stalwart-migrate/     main.go, preflight.go, run.go, status.go — CLI entry + wiring
  internal/plan/            version-boundary → ordered phase list (§4.6)                 [done]
  internal/checkpoint/      run-id, state.json read/write, resume logic (§5)             [done]
  internal/preflight/       §4.1 checks                                                  [done]
  internal/backup/          §4.2 — fs/db snapshot, settings dump+convert, Vandelay export [done]
  internal/recovery/        §4.4 — recovery-mode process supervision + apply             [done]
  internal/cutover/         §4.5 — binary swap, unit rewrite, restart, quota rebuild    [done, unwired]
  internal/validate/        §4.7 — boot-check + content-integrity done; DKIM/TLS + mail-flow not yet [partial]
  internal/service/         systemd/Docker start+stop, used by §4.5                      [done]
  internal/stalwartapi/     Ping, AccountSnapshot (per-mailbox counts via impersonation), quota tasks [done]
  internal/config/          tool's own config (paths, thresholds, credentials handling)  [not started]
  docs/                     this file + phase-specific notes as they get built out

There's no separate internal/stage package: the convert half of migrate_v016.py lives in internal/backup next to dump (same script, same invocation pattern). The sandbox-cloning logic that used to stand in for the rest of §4.3 lived directly in cmd/run.go and goes away with the sandbox (§4.9); what §4.3 still needs is the apply-plan generator, which has no code yet at all.

There's no internal/rollback either, and that's a deliberate removal rather than a gap — see §4.8.

internal/stalwartapi is deliberately the only thing that speaks JMAP/HTTP to Stalwart — every other package depends on it, not on net/http directly, so auth handling and retry/backoff live in one place. internal/service is the same idea for the other external surface: it is the only thing that shells out to systemctl or docker, so the commands that can take mail delivery down sit in one auditable file rather than in each phase that happens to need them. preflight.DeploymentKind is a type alias for service.Kind, so detection and control can't drift apart.

8. Open questions for the next pass

• Credential handling: recovery-mode admin password and any stored JMAP credentials need a real secrets story (env var pass-through is fine for v1, but the checkpoint file must never contain them in plaintext).

• Cluster orchestration: §4.1's cluster gate assumes the operator stops other nodes manually; a v2 could SSH-coordinate that instead. Out of scope for v1.

• migrate_v016.py dependency: pinning by hash is a start, but the script is Stalwart's, not ours — need a policy for what happens when it changes upstream (re-vendor + re-test before bumping the pin, never silently float to main).

• Most of the "unmigrated" settings are not work at all - measured, then classified. The raw figure from a production instance was 12,182 settings, which reads as an impossible amount of manual reconstruction. Snapshotting a migrated v0.16.14 store showed why it is misleading: 8,547 of them are server.blocked-ip, runtime auto-ban state that repopulates itself; 3,337 are stock spam-filter and lookup data v0.16 ships its own copies of (2,084 MemoryLookupKey, 66 SpamRule, 18 SpamDnsblServer were already present after migration); ~224 had already been carried by another route, DKIM included, as DkimSignature objects with their private keys. That leaves ~293 keys genuinely needing a human.

  backup.UnmigratedReport.Classify encodes this, and the rehearsal reports the categorised view rather than the raw count. Every rule was checked against a migrated instance rather than inferred, and an unrecognized prefix defaults to "needs review" - assuming an unknown setting is safe to ignore is the wrong default.

  This also retired a planned pair of generators. v0.15's spam rules are stwt_rbl_senderscore_ip; v0.16's are STWT_RBL_SENDERSCORE_IP - the same stock set, already installed. Generating them from v0.15 would duplicate every rule and revert a year of upstream updates, so the lookup and spam-filter generators were deliberately not written. The targets worth generating are the small site-specific groups: queue.schedule, queue.tls, session.auth, server.auto-ban.

• Settings apply-plan (§4.3): started, and the critical path. internal/applyplan covers server.listener — verified end to end by applying a generated plan to a real 0.16.14 instance and reading back all ten listeners with correct protocols, binds and TLS flags. Everything else in the worklist is still manual: the largest groups are lookup.url-redirectors, lookup.trusted-domains, spam-filter.list, spam-filter.rule and spam-filter.dnsbl, each needing its own confirmed mapping (x:StoreLookup, x:HttpLookup, x:SpamRule, x:SpamDnsblServer). The generated plan should still require explicit operator sign-off even with --yes set for everything else.

• Account/mailbox enumeration (stalwartapi.Client.AccountSnapshot): implemented, including per-mailbox message counts. Account count and domains come from x:Account/query + x:Account/get against Stalwart's management API (/api, capability urn:stalwart:jmap), confirmed against crates/jmap/src/principal/{get,query}.rs and docs/ref/object/account.md. Per-mailbox counts needed a second research pass, because a superuser's own JMAP session does not implicitly grant cross-account access — confirmed by reading crates/jmap/src/api/session.rs: the session's accounts map is built solely from the authenticated identity's own membership/sharing grants, unaffected by any admin flag. The real, documented mechanism is Stalwart's impersonate permission (docs/auth/authorization/administrator.md): an account holding it can log in as another account via the composite Basic-auth username <target>%<impersonator>, after which standard RFC 8621 Mailbox/get (property totalEmails) works normally against that impersonated session's own apiUrl (session-discovered per RFC 8620, not the /api management endpoint — confirmed as a distinct endpoint in docs/ref/object/account.md). AccountSnapshot now does this per account it finds; a single account's failure (most likely: impersonate not granted) is recorded in Snapshot.MailboxErrors rather than failing the whole snapshot, so one misconfigured account doesn't hide a working result for every other one. One resolved false alarm worth recording: an initial pass of this same research, reading Stalwart's main branch source directly, reported x:Account apparently replaced by x:Principal/x:Quota. Checking the published docs site directly (which has no principal.md/quota.md page, and still documents x:Account with a working example) showed that was an unreleased/in-development refactor in main, not the interface the current released version actually exposes — a reminder that "read the source" and "read what's actually shipped" can disagree, and it's worth checking both before changing already-working code on the strength of one. Superseded in part. That mailbox-count comparison was verified only against fabricated fixtures, and against a real 0.15.5 source it does not work at all: 0.15.x reports no per-mailbox counts and refuses the impersonation login, so the "before" side is always empty, and the comparison used to iterate that empty map and report "all message counts match" — a vacuous pass on the strongest claim this tool makes. It now states plainly when counts were not compared (§4.7). Account and domain enumeration against 0.15.x works via its REST principal API. Still open: preflight/validate attempt every account serially with no sampling, which could be slow on a large install; and DKIM/TLS fingerprint checks plus a live mail-flow SMTP→IMAP smoke test (the rest of §4.7's suite) aren't implemented.

• Cutover is built; nothing wires it into a production run yet. internal/cutover (§4.5) and internal/service are implemented and tested against fakes. run still refuses, for one remaining reason: §4.3 stage doesn't exist, and neither does the production pipeline that would run preflight → backup → stage → recovery-mode → cutover → validate against real paths. What stage still needs: downloading and verifying the target binary into a staging path (preflight.ResolveRelease and backup.DownloadFile between them already have the pieces), running convert against real paths, and the settings apply-plan — which, per the measurement above, is what decides whether the migrated server serves anything at all.

• What has and hasn't been proven against real software. A smoke VM (Debian 13, real Stalwart 0.15.5 under a real systemd unit, RocksDB, seeded accounts and mail) has now exercised preflight, backup, the settings dump, migrate_v016.py convert, and the recovery-mode store migration end to end - and a scrubbed copy of a production settings corpus has been through the converter. Everything in §4.9's rewrite and most of §8's newer entries came from that, not from reading code.

  Cutover has now run, in a complete 0.15.5 -> 0.16.14 migration of the smoke VM: binary verified and installed, the unit rewritten with its hardening intact, service restarted, health check passed, and checkpoint resume exercised. All mail survived and was readable afterwards. Three defects came out of it and are fixed:

  • The converted config was installed root-owned while the service runs as its own user, so it crash-looped 28 times on "Permission denied". The ownership trap that retired the rollback implementation (§4.8), in a new place. Cutover now installs the config itself, copying ownership and mode from the config being replaced.
  • v0.16.14 does not serve /api. Confirmed against a fully migrated, fully configured instance, not just a sandbox: /api, /api/principal and /jmap/ all 404, and the JMAP endpoint is the one the session document advertises. The client now discovers it, re-basing the advertised path onto the operator's own host - a real instance advertises a canonical public URL that frequently isn't reachable from where this tool runs.
  • Dispatching on the urn:stalwart:jmap capability was wrong, because neither version advertises it. The client probes what the instance actually serves instead.

  Still unproven: the x:Task quota wire format (the endpoint fix makes it reachable, but the migrated instance refused the call - below), systemd drop-in handling, and anything on a non-RocksDB backend or a Docker deployment.

• Rehearsal has been run against a live production instance, read-only, and found two preflight defects a test instance could not have:

  • Store-backend detection missed a flat config entirely. A config generated by stalwart --init declares type = "rocksdb" inside a [store.rocksdb] section; the production config has no section headers at all and declares store.rocksdb.type = "rocksdb" flat. Detection checked only for a bare type key, so it found nothing and left topology.store_backend empty - and backup.Run treated an unrecognized backend as a skip, meaning a real run would have proceeded with no filesystem backup whatsoever. Both fixed: flat keys are detected, and an unrecognized backend is now a hard failure, because the one artifact this phase exists to produce must not be quietly absent.
  • The cluster warning didn't say where it matched. It fires on any occurrence of "cluster" anywhere in the config, which is the right bias
    • a missed cluster is the dangerous direction - but on the production config the sole match was inside the value of an unrelated setting. The warning now names the location and says whether the match was the setting or only its value, which turns a config-wide search into a glance.

  Also confirmed there: the HTTPS path works against a real certificate, and the coverage numbers match the earlier scrubbed-corpus measurement exactly (12,182 unmigrated). Account enumeration past one page is still untested - that instance has six accounts, not the hundred-plus the pagination loop exists for.

• A migrated instance has no working administrator - diagnosed and fixed. migrate_v016.py assigns every migrated account the User role regardless of what it held before, so an account that was an administrator in v0.15 comes out authenticating normally and refused every management call. Ordinary users are unaffected: User is what they had and what they get, and their credentials, mail and mailboxes all survive untouched.

  The v0.16 shape came from the server's own schema document (GET /api/schema): x:UserRoles is a multi-variant type with variants User, Admin and Custom, and Account is itself multi-variant, so the upsert needs its own @type as well. applyplan.AccountRoleOperations restores it from the principals dump, emitting operations only for accounts whose role actually changes - rewriting every account would be a far larger blast radius for no benefit. Where v0.15 listed several roles, admin wins and the collapse is reported; roles with no v0.16 equivalent are named rather than dropped silently.

• x:Account.domainId returns an internal id on v0.16 - resolved. A pre-migration snapshot records domains as names ("smoke.test"); the same instance afterwards reported "b", so the §4.7 directory comparison would have read every domain as having vanished. The client now resolves ids to names with x:Domain/query + x:Domain/get in a single request via a JMAP back-reference, confirmed against a live 0.16.14. An id that cannot be resolved is kept as-is rather than dropped, since a domain that can't be named is still a domain that exists; a failure of the resolution call itself is an error, because silently comparing ids against names is the bug this fixes.

• Quota recalculation is grounded but unproven. The x:Task wire format comes from Stalwart's schema reference rather than a live server; §4.5 lists exactly which two details are inferred. A smoke test against a real 0.16 instance would settle both, and would let this step be promoted from "warns on failure" to a hard check.

• Docker is wired end to end and has never met a real Stalwart image. Preflight inspects a container and blocks on what stands in the way; stage pulls and verifies an image; the recovery cycle runs in a throwaway container against the live data; cutover recreates the container, refusing one whose definition it would not carry across intact. Every test drives a fake docker, which proves the right commands are assembled and proves nothing about whether the image reads the config it is handed - the same limit §4.8 records about the deleted rollback code, and the reason run refuses a container without --container-path-unproven. A rehearsal on a clone, then a real migration, is what would retire that flag. Compose stays refused deliberately: recreating a compose-managed container out from under compose leaves the container and the compose file disagreeing, and the next compose up reverts the migration.

• The converted config reaches a container through the data volume. run writes it under the host side of whichever mount covers --data-dir and names it on the container side, because cutover recreates a container with the mounts it had and cannot invent a new one for a config file. That is an inference from how the mounts must line up rather than something a real deployment has confirmed.

• Cutover ignores systemd drop-ins. It rewrites only the main unit file, so an ExecStart or Environment override in /etc/systemd/system/stalwart.service.d/*.conf is invisible to it - including a recovery-mode variable set there, which is exactly the footgun the rewrite exists to prevent. Drop-ins are common enough that this needs handling before a production run, at minimum by detecting them and refusing.

• Nothing prevents concurrent runs. Two invocations against the same run-id would both proceed; there's no lock file or equivalent.

• A full dress rehearsal has been run against a clone of production - the real 3.6 GB store, 12,361 settings, 6 accounts across 9 domains, streamed into the smoke VM and migrated 0.15.5 -> 0.16.14 with the tool's own phases. It succeeded, and the timings are the useful part: the recovery-mode conversion of that store took 2 seconds, and the whole sequence from service-stop to service-start was seconds of work. A migration window is dominated by waiting and verification, not by data volume - worth knowing before scheduling one around store size.

  Four things it found that smaller instances could not:

  • Account roles broke on production-shaped names. v0.16 stores an account as a local part plus a domain reference; the generator was passing v0.15's full address and the server rejected it ("Invalid email local part"). The smoke instance used bare usernames and never exercised it. Fixed - and because local parts are unique only within a domain, an ambiguous one is now refused with a warning rather than risking Admin landing on the wrong account.
  • A failed apply leaves the store in bootstrap mode. After a partial apply the instance answers every management call with "The server is in bootstrap mode. Only the 'Bootstrap' object type can be accessed until the bootstrap process is complete." So a half-applied plan is not a partially configured server, it is an unusable one, which raises the stakes on apply failures considerably.
  • A config fallback-admin does not survive the migration. v0.16's config is a store pointer, so an [authentication.fallback-admin] block in the old config.toml simply ceases to exist. The credentials an operator supplies for the pre-migration instance therefore stop working on the migrated one, and cutover's health check - which authenticated - failed a cutover that had actually succeeded. Liveness and credentials are now separate questions: any response proves the service is up, and credentials that no longer work are a warning naming this cause.
  • tenant-admin has no v0.16 equivalent and is reported as unrestorable rather than silently dropped.

• A domain and the accounts on it must agree about their tenant in v0.16, and Stalwart's converter does not make them agree. A second live attempt on 2026-08-24 got further - preflight clean, binary staged, settings dumped and converted - and then failed during recovery-mode migration, again with the service already stopped:

  created Tenant (1)
  created Domain (9)
  create Account restore-13: invalidForeignKey | Object id: Domain#d
  

  In v0.15 a domain's tenant and a principal's tenant were independent facts. v0.16 requires a tenant-scoped Account to sit on a Domain owned by that same tenant - for its primary domain and for every alias - and answers invalidForeignKey on the Domain reference otherwise. migrate_v016.py carries the two facts over independently: _build_domains sets a domain's memberTenantId only when the domain appears as a declared domain principal carrying a tenant, while _build_user sets the account's from the account's own record. A domain that exists only inside an email address is inferred, gets no tenant, and any tenant-scoped account using it is then rejected.

  This was established by reproduction, not inference. A synthetic v0.15 principal dump run through the unpatched converter and applied to a real 0.16.14 in recovery mode reproduces invalidForeignKey | Object id: Domain#d character for character - the #d is the server's own object id for the offending domain, not a plan client-id. The same harness shows which directions are actually constrained:

  account	domain	result
  tenant-scoped	no tenant	rejected
  tenant-scoped	a different tenant	rejected
  tenant-scoped	its own tenant	accepted
  global	tenant-owned	accepted

  Only the first two fail, and only the first is repairable: where a tenant-less domain is used exclusively by accounts of one tenant, giving the domain that tenant is the sole assignment that both applies and keeps every account. applyplan.ReconcileDomainTenants does that to the plan between convert and apply, reports each adoption, and refuses - without modifying anything - when the accounts genuinely disagree, since forcing such a plan through would mean dropping mailboxes. The same fix has been prepared for migrate_v016.py upstream; the tool downloads that script rather than vendoring it, so the repair lives here until a released version carries it, and is a no-op on a plan that is already consistent.

  An earlier version of this section claimed the converter emitted every Account with tenantId: null. That was wrong: the field is memberTenantId, the converter does populate it, and the export had been inspected for a key no version of the script ever writes. Preflight briefly refused every multi-tenant install on the strength of that misreading. The refusal is now narrowed to the arrangements v0.16 truly cannot represent.

  The tool's failure was in when this was discovered. Tenant membership is readable while the server is still running, so preflight now maps it (stalwartapi.Client.FetchTenantLayout), predicts the outcome with the same rule the server enforces, and either warns about the domains that will adopt a tenant or fails - before anything is touched. That is the same lesson as the stalwart-cli check immediately below: both were knowable in advance, and both were found after a production mail server had been stopped. Any future dependency of the conversion belongs in preflight, not in the phase that needs it.

• A live migration attempt failed and cost a restore. Three defects, all fixed, all now proven against a reproduction. On 2026-08-24 a real migration stopped a production mail server and then discovered the host's stalwart-cli was 0.13.4 - present, but from when the CLI shipped with the server, and with no apply command. Recovery was closed in both directions: v0.16's recovery-mode boot had already bumped the store schema to v6 (expected 5 or below, found 6), so the old binary could not reopen it, and going forward needed export.json, which the failure path had deleted - regenerating it required a settings dump from a live v0.15 instance that could no longer start. The operator restored a day-old snapshot and lost a day of mail.

  • Preflight now verifies the external tools (CheckExternalTools), before anything is touched: stalwart-cli must exist and be v1.0.2 or later, and python3 must run. Every fact needed to prevent that outage was available in under a second from a stopped state. Skipped entirely for a patch upgrade, which invokes neither tool.
  • A failed run no longer deletes its own inputs. The cleanup applied to every exit path, which was right for a sandboxed rehearsal and catastrophic here: after the service is stopped, the settings dump cannot be regenerated, so deleting it removes the only way forward.
  • run --resume <id> continues an interrupted run. The checkpoint machinery existed but never engaged, because run created a new run every invocation - so a retry would re-run preflight against a binary already moved aside and fail. Completed steps are now skipped.

  Verified on a VM built to match: stalwart-cli 0.15.5 installed, accounts and mail seeded. Preflight refused with the service still running and mail still flowing; a stub CLI that passed the version check and failed the apply left the run stopped with all eight inputs intact; and --resume carried it to a clean finish - five seconds of downtime, quotas rebuilt. That failure-path test is the one that should have run before production, and did not.

• run is built and works. preflight -> stage -> dump -> preserve -> stop -> convert -> supplement -> recovery-mode -> cutover, checkpointed throughout, verified end to end against a real 0.15.5: mail down for six seconds, users unchanged, and recalculate-quotas succeeding for the first time - the x:Task wire format inferred from the schema reference turned out to be right, once the endpoint discovery and role restoration made it reachable at all.

  Two gates, deliberately separate: --yes is about intent, and --recovery-point-confirmed is a claim about the world that this tool cannot verify and must not assume. internal/stage (§4.3) fetches the release, refuses to substitute a different build for the one it wants, honours a pinned checksum, and confirms the extracted binary reports the version its tag claims - because everything upstream of that check is an assumption about somebody else's release process.

• rehearse (§4.9) is designed but not built. The command is still run --dry-run with the old sandbox-cloning shape. Building it is mostly deletion: the dump, convert and report pieces already exist and work against real instances; what goes away is the backup clone, the sandbox, the recovery-mode cycle and the boot check. cmd/run.go's sandbox logic disappears with it, which also removes the reason §7 gives for there being no internal/stage package.

• Post-migration validation has no reachable instance to validate (§4.7/§4.9). Until the apply-plan reconstructs listeners, nothing that boots from a converted store can answer an API, so "did the migration preserve the data" cannot be asked of the migrated instance at all. This is the strongest argument for building the apply-plan first.

Sources

Grounded in Stalwart's own documentation and community reports as of 2026-08-19:

• UPGRADING/v0_16.md — exact migration procedure this tool automates
• UPGRADING/v0_15.md — prior breaking-change boundary
• Database Migration docs
• Backup docs (Vandelay)
• Upgrading guide
• v0.16 blog post
• Discussion #2892 — breaking changes overview
• Discussion #3004 — upgrading Q&A
• Discussion #3025 — real post-migration WebUI login failure
• CHANGELOG.md — confirms 0.16.1–0.16.14 carry no further schema migrations