Files
stack/packages/storage/src/migrate-tier.ts
jason.woltje 15d849c166
Some checks failed
ci/woodpecker/push/ci Pipeline was successful
ci/woodpecker/push/publish Pipeline failed
test(storage): integration test for migrate-tier (FED-M1-08) + camelCase column fix (#477)
2026-04-20 01:40:02 +00:00

715 lines
24 KiB
TypeScript

/**
* migrate-tier.ts — Core logic for `mosaic storage migrate-tier`.
*
* Migrates data from `tier: local` (PGlite, normalized Drizzle schema) or
* `tier: standalone` (Postgres without pgvector) → `tier: federated`
* (Postgres + pgvector).
*
* Source: DrizzleMigrationSource — reads from the NORMALIZED Drizzle/relational
* schema tables (not the flat `id TEXT + data JSONB` PgliteAdapter schema).
* Both local (PGlite) and standalone (Postgres) sources use the same Drizzle
* abstraction via createPgliteDb() or createDb() from @mosaicstack/db.
* Target: PostgresMigrationTarget — upserts via raw SQL into the same schema.
*
* Key design decisions:
* - Tables are migrated in topological (FK-dependency) order so that
* parent rows exist before child rows are inserted.
* - sessions + verifications are skipped — they are ephemeral / TTL'd.
* - adminTokens is skipped — token hashes are environment-specific
* and should be re-issued on the target.
* - insights.embedding is omitted from source SELECT when the source lacks
* pgvector (local/PGlite tier); target insert gets NULL for that column.
* insights.embedding is nullable per schema (no .notNull() constraint).
* - Each table's batch is wrapped in a transaction for atomicity.
* - Upsert semantics (ON CONFLICT DO UPDATE) make re-runs idempotent.
*
* TODO (FED-M1-08): Add integration tests against real PGlite → real PG.
*/
import postgres from 'postgres';
import * as schema from '@mosaicstack/db';
import { sql as drizzleSql } from '@mosaicstack/db';
/* ------------------------------------------------------------------ */
/* Types */
/* ------------------------------------------------------------------ */
export interface MigrationSource {
/**
* Return all rows from a table (normalized Drizzle schema rows).
* When sourceHasVector is false and the table has a vector column,
* the source MUST omit the vector column from the result and the
* caller will set it to null (see normaliseSourceRow).
*/
readTable(
tableName: string,
opts?: { limit?: number; offset?: number },
): Promise<Record<string, unknown>[]>;
/** Count rows in a table. */
count(tableName: string): Promise<number>;
/** Close the source connection. */
close(): Promise<void>;
}
export interface MigrationTarget {
/**
* Upsert a batch of rows into a table.
* Must use ON CONFLICT (id) DO UPDATE semantics.
*/
upsertBatch(tableName: string, rows: Record<string, unknown>[]): Promise<void>;
/**
* Count rows in a target table.
*/
count(tableName: string): Promise<number>;
/**
* Check whether pgvector extension is installed.
*/
hasPgvector(): Promise<boolean>;
/** Close the target connection. */
close(): Promise<void>;
}
export interface MigrateTierOptions {
/** Target postgres connection URL. */
targetUrl: string;
/** Whether to skip all writes (dry-run). */
dryRun: boolean;
/** Skip the non-empty target guard. */
allowNonEmpty: boolean;
/** Rows per transaction batch. */
batchSize: number;
/** Called with progress messages. */
onProgress: (msg: string) => void;
}
export interface TableMigrationResult {
table: string;
rowsMigrated: number;
skipped: boolean;
skipReason?: string;
}
export interface MigrateTierResult {
tables: TableMigrationResult[];
totalRows: number;
dryRun: boolean;
}
/* ------------------------------------------------------------------ */
/* Schema: FK-aware topological table order */
/* ------------------------------------------------------------------ */
/**
* SKIP_TABLES: ephemeral or environment-specific tables not worth migrating.
*
* - sessions: TTL'd auth sessions — invalid in new environment.
* - verifications: one-time tokens (email verify, etc.) — already expired.
* - admin_tokens: hashed tokens bound to old environment keys — re-issue.
*/
export const SKIP_TABLES = new Set(['sessions', 'verifications', 'admin_tokens']);
/**
* Topologically ordered table list (parents before children).
*
* Derived from FK references in packages/db/src/schema.ts:
*
* users (no FKs)
* teams → users
* team_members → teams, users
* accounts → users
* projects → users, teams
* agents → projects, users
* missions → projects, users
* tasks → projects, missions
* mission_tasks → missions, tasks, users
* conversations → users, projects, agents
* messages → conversations
* preferences → users
* insights → users [has embedding vector column]
* agent_logs → users
* skills → users (installedBy, nullable)
* routing_rules → users (userId, nullable)
* provider_credentials → users
* appreciations (no FKs)
* events (no FKs)
* tickets (no FKs)
* summarization_jobs (no FKs)
*
* Skipped (not in this list):
* sessions → users (ephemeral)
* verifications (no FKs, ephemeral)
* admin_tokens → users (environment-specific)
*/
export const MIGRATION_ORDER: string[] = [
'users',
'teams',
'team_members',
'accounts',
'projects',
'agents',
'missions',
'tasks',
'mission_tasks',
'conversations',
'messages',
'preferences',
'insights',
'agent_logs',
'skills',
'routing_rules',
'provider_credentials',
'appreciations',
'events',
'tickets',
'summarization_jobs',
];
/** Tables that carry a vector embedding column on the target. */
const VECTOR_TABLES = new Set(['insights']);
/* ------------------------------------------------------------------ */
/* Utility: derive topological order from an adjacency list */
/* ------------------------------------------------------------------ */
/**
* Given an adjacency list (table → list of tables it depends on),
* return a valid topological ordering (Kahn's algorithm).
*
* Exposed for unit testing.
*/
export function topoSort(deps: Map<string, string[]>): string[] {
const nodes = [...deps.keys()];
const inDegree = new Map<string, number>();
const adjReverse = new Map<string, string[]>();
for (const node of nodes) {
if (!inDegree.has(node)) inDegree.set(node, 0);
if (!adjReverse.has(node)) adjReverse.set(node, []);
for (const dep of deps.get(node) ?? []) {
inDegree.set(node, (inDegree.get(node) ?? 0) + 1);
if (!adjReverse.has(dep)) adjReverse.set(dep, []);
adjReverse.get(dep)!.push(node);
}
}
// Start with nodes that have no dependencies.
const queue: string[] = [];
for (const [node, deg] of inDegree) {
if (deg === 0) queue.push(node);
}
const result: string[] = [];
while (queue.length > 0) {
const node = queue.shift()!;
result.push(node);
for (const dependent of adjReverse.get(node) ?? []) {
const newDeg = (inDegree.get(dependent) ?? 0) - 1;
inDegree.set(dependent, newDeg);
if (newDeg === 0) queue.push(dependent);
}
}
if (result.length !== nodes.length) {
throw new Error('Cycle detected in FK dependency graph');
}
return result;
}
/**
* Return the migration table order, excluding SKIP_TABLES.
* Uses the pre-computed MIGRATION_ORDER (verified against schema.ts).
*/
export function getMigrationOrder(): string[] {
return MIGRATION_ORDER.filter((t) => !SKIP_TABLES.has(t));
}
/* ------------------------------------------------------------------ */
/* TABLE_OBJECTS: migration table name → Drizzle table object */
/* ------------------------------------------------------------------ */
/**
* Maps MIGRATION_ORDER table names to their corresponding Drizzle table
* objects from the normalized schema. Used by DrizzleMigrationSource to
* execute typed `db.select().from(table)` queries.
*
* Keyed by snake_case table name (matching MIGRATION_ORDER + SKIP_TABLES).
*/
// eslint-disable-next-line @typescript-eslint/no-explicit-any
const TABLE_OBJECTS: Record<string, any> = {
users: schema.users,
teams: schema.teams,
team_members: schema.teamMembers,
accounts: schema.accounts,
projects: schema.projects,
agents: schema.agents,
missions: schema.missions,
tasks: schema.tasks,
mission_tasks: schema.missionTasks,
conversations: schema.conversations,
messages: schema.messages,
preferences: schema.preferences,
insights: schema.insights,
agent_logs: schema.agentLogs,
skills: schema.skills,
routing_rules: schema.routingRules,
provider_credentials: schema.providerCredentials,
appreciations: schema.appreciations,
events: schema.events,
tickets: schema.tickets,
summarization_jobs: schema.summarizationJobs,
// Skipped tables — included so count() works for preflight but never passed
// to upsertBatch.
sessions: schema.sessions,
verifications: schema.verifications,
admin_tokens: schema.adminTokens,
};
/* ------------------------------------------------------------------ */
/* DrizzleMigrationSource */
/* ------------------------------------------------------------------ */
/**
* MigrationSource backed by a Drizzle DB handle (works with both
* PostgresJsDatabase and PgliteDatabase — they share the same Drizzle
* query API for schema-defined tables).
*
* For the `insights` table (the only vector-column table), when the source
* lacks pgvector (local/PGlite tier), the `embedding` column is excluded
* from the SELECT projection via a raw `db.execute()` query that lists
* only non-vector columns. This prevents a type-registration error from
* PGlite, which does not know the `vector` type. The caller (runMigrateTier
* via normaliseSourceRow) will set embedding to null on the resulting rows.
*
* Column projection is opt-in: pass `sourceHasVector: false` to activate it.
*/
export class DrizzleMigrationSource implements MigrationSource {
constructor(
// eslint-disable-next-line @typescript-eslint/no-explicit-any
private readonly db: any,
private readonly sourceHasVector: boolean = true,
) {}
/**
* Columns of the insights table that do NOT include the vector embedding.
* Used for the no-pgvector projection path.
*/
private static readonly INSIGHTS_COLUMNS_NO_VECTOR = [
'id',
'user_id',
'content',
'source',
'category',
'relevance_score',
'metadata',
'created_at',
'updated_at',
'decayed_at',
] as const;
async readTable(
tableName: string,
opts?: { limit?: number; offset?: number },
): Promise<Record<string, unknown>[]> {
const table = TABLE_OBJECTS[tableName];
if (!table) throw new Error(`DrizzleMigrationSource: unknown table "${tableName}"`);
// For vector tables when source lacks pgvector: use column-allowlist raw query
// to avoid type-registration errors.
if (VECTOR_TABLES.has(tableName) && !this.sourceHasVector) {
const cols = DrizzleMigrationSource.INSIGHTS_COLUMNS_NO_VECTOR.map((c) => `"${c}"`).join(
', ',
);
let sql = `SELECT ${cols} FROM "${tableName}"`;
const params: unknown[] = [];
if (opts?.limit !== undefined) {
params.push(opts.limit);
sql += ` LIMIT $${params.length.toString()}`;
}
if (opts?.offset !== undefined) {
params.push(opts.offset);
sql += ` OFFSET $${params.length.toString()}`;
}
// eslint-disable-next-line @typescript-eslint/no-explicit-any
const result = await (this.db as any).execute(
// drizzle-orm/pglite and drizzle-orm/postgres-js both accept a raw
// SQL template; use the tagged-template sql helper from drizzle-orm.
// Since we need dynamic params, we use db.execute with a raw string
// via the PGlite/postgres.js driver directly.
{ sql, params, typings: [] },
);
// drizzle execute returns { rows: unknown[][] } for PGlite driver,
// or a RowList for postgres.js. Normalise both shapes.
// eslint-disable-next-line @typescript-eslint/no-explicit-any
const raw = result as any;
if (Array.isArray(raw)) {
// postgres.js shape: array of row objects
return raw as Record<string, unknown>[];
}
if (raw?.rows && Array.isArray(raw.rows)) {
// PGlite shape: { rows: unknown[][] } OR { rows: Record<string,unknown>[] }
const rows = raw.rows as unknown[];
if (rows.length === 0) return [];
if (Array.isArray(rows[0])) {
// Columnar: convert to objects using fields array if available
const fields: string[] =
(raw.fields as Array<{ name: string }> | undefined)?.map((f) => f.name) ??
DrizzleMigrationSource.INSIGHTS_COLUMNS_NO_VECTOR.slice();
return (rows as unknown[][]).map((row) => {
const obj: Record<string, unknown> = {};
for (let i = 0; i < fields.length; i++) {
obj[fields[i]!] = row[i];
}
return obj;
});
}
return rows as Record<string, unknown>[];
}
return [];
}
// Standard Drizzle select for all other tables (and vector tables when
// the source has pgvector registered).
// eslint-disable-next-line @typescript-eslint/no-explicit-any
let query = (this.db as any).select().from(table);
if (opts?.limit !== undefined) query = query.limit(opts.limit);
if (opts?.offset !== undefined) query = query.offset(opts.offset);
return (await query) as Record<string, unknown>[];
}
async count(tableName: string): Promise<number> {
const table = TABLE_OBJECTS[tableName];
if (!table) throw new Error(`DrizzleMigrationSource: unknown table "${tableName}"`);
// eslint-disable-next-line @typescript-eslint/no-explicit-any
const [row] = await (this.db as any)
.select({ n: drizzleSql<number>`COUNT(*)::int` })
.from(table);
return (row as { n: number } | undefined)?.n ?? 0;
}
async close(): Promise<void> {
// Lifecycle managed externally — caller closes the db handle.
}
}
/* ------------------------------------------------------------------ */
/* Real postgres target adapter */
/* ------------------------------------------------------------------ */
/**
* Live implementation of MigrationTarget backed by a real Postgres connection.
* Used by the CLI; mocked in tests.
*/
export class PostgresMigrationTarget implements MigrationTarget {
private sql: ReturnType<typeof postgres>;
constructor(url: string) {
this.sql = postgres(url, {
max: 5,
connect_timeout: 10,
idle_timeout: 30,
});
}
async upsertBatch(tableName: string, rows: Record<string, unknown>[]): Promise<void> {
if (rows.length === 0) return;
// Collect all column names from the batch (union of all row keys).
const colSet = new Set<string>();
for (const row of rows) {
for (const k of Object.keys(row)) colSet.add(k);
}
const cols = [...colSet];
if (!cols.includes('id')) {
throw new Error(`Table ${tableName}: rows missing 'id' column`);
}
// Build VALUES list — use postgres tagged-template helpers for safety.
// postgres.js supports bulk inserts via array of objects.
// eslint-disable-next-line @typescript-eslint/no-explicit-any
await this.sql.begin(async (tx: any) => {
// Insert in chunks to avoid enormous single queries.
for (let i = 0; i < rows.length; i += 500) {
const chunk = rows.slice(i, i + 500);
// Normalise rows: fill missing columns with null.
const normalised = chunk.map((row) => {
const out: Record<string, unknown> = {};
for (const col of cols) {
out[col] = row[col] ?? null;
}
return out;
});
const colList = cols.map((c) => `"${c}"`).join(', ');
const updateList = cols
.filter((c) => c !== 'id')
.map((c) => `"${c}" = EXCLUDED."${c}"`)
.join(', ');
// Build values placeholders
const valuePlaceholders = normalised
.map((_, ri) => `(${cols.map((_, ci) => `$${ri * cols.length + ci + 1}`).join(', ')})`)
.join(', ');
const flatValues = normalised.flatMap((row) => cols.map((c) => row[c] ?? null));
const query = `
INSERT INTO "${tableName}" (${colList})
VALUES ${valuePlaceholders}
ON CONFLICT (id) DO UPDATE SET ${updateList}
`;
await tx.unsafe(query, flatValues as never[]);
}
});
}
async count(tableName: string): Promise<number> {
const rows = await this.sql.unsafe(`SELECT COUNT(*)::int AS n FROM "${tableName}"`);
const row = rows[0] as unknown as { n: number } | undefined;
return row?.n ?? 0;
}
async hasPgvector(): Promise<boolean> {
const rows = await this.sql`
SELECT 1 FROM pg_extension WHERE extname = 'vector'
`;
return rows.length > 0;
}
async close(): Promise<void> {
await this.sql.end();
}
}
/* ------------------------------------------------------------------ */
/* Source-row normalisation */
/* ------------------------------------------------------------------ */
/**
* Convert a camelCase key to snake_case.
* e.g. "userId" → "user_id", "emailVerified" → "email_verified".
* Keys that are already snake_case (no uppercase letters) are returned as-is.
*/
function toSnakeCase(key: string): string {
return key.replace(/[A-Z]/g, (c) => `_${c.toLowerCase()}`);
}
/**
* Drizzle returns rows as camelCase TypeScript objects (e.g. `userId`, not
* `user_id`). The PostgresMigrationTarget upserts via raw SQL and uses the
* column names as given. We must convert camelCase keys → snake_case before
* building the INSERT statement so column names match the PG schema.
*
* Exception: the `insights` no-vector path already returns snake_case keys
* from its raw SQL projection — toSnakeCase() is idempotent for already-
* snake_case keys so this conversion is safe in all paths.
*
* For vector tables (insights), if `embedding` is absent from the source row
* (because DrizzleMigrationSource omitted it in the no-vector projection), we
* explicitly set it to null so the target ON CONFLICT UPDATE doesn't error.
*
* NOTE: insights.embedding is defined as `vector('embedding', { dimensions:
* 1536 })` with no `.notNull()` in schema.ts — it accepts NULL.
*/
export function normaliseSourceRow(
tableName: string,
row: Record<string, unknown>,
sourceHasVector: boolean,
): Record<string, unknown> {
// Convert all camelCase keys to snake_case for raw-SQL target compatibility.
const out: Record<string, unknown> = {};
for (const [k, v] of Object.entries(row)) {
out[toSnakeCase(k)] = v;
}
if (VECTOR_TABLES.has(tableName) && !sourceHasVector) {
// Source cannot have embeddings — explicitly null them so ON CONFLICT
// UPDATE doesn't try to write undefined.
out['embedding'] = null;
}
return out;
}
/* ------------------------------------------------------------------ */
/* Precondition checks */
/* ------------------------------------------------------------------ */
export class MigrationPreconditionError extends Error {
constructor(
message: string,
public readonly remediation: string,
) {
super(message);
this.name = 'MigrationPreconditionError';
}
}
/**
* Verify target preconditions before writing any data.
*
* Checks:
* 1. pgvector extension installed.
* 2. User-data tables are empty (unless --allow-non-empty).
*/
export async function checkTargetPreconditions(
target: MigrationTarget,
allowNonEmpty: boolean,
tablesToMigrate: string[],
): Promise<void> {
const hasVector = await target.hasPgvector();
if (!hasVector) {
throw new MigrationPreconditionError(
'Target Postgres does not have the pgvector extension installed.',
'Run: CREATE EXTENSION IF NOT EXISTS vector; — or use the pgvector/pgvector:pg17 Docker image.',
);
}
if (!allowNonEmpty) {
// Check the first non-empty user-data table.
for (const table of tablesToMigrate) {
const n = await target.count(table);
if (n > 0) {
throw new MigrationPreconditionError(
`Target table "${table}" already contains ${n.toString()} rows.`,
'Pass --allow-non-empty to overwrite existing data (upsert semantics), ' +
'or point to an empty target database.',
);
}
}
}
}
/* ------------------------------------------------------------------ */
/* Core migration runner */
/* ------------------------------------------------------------------ */
/**
* Run the tier migration.
*
* @param source Adapter for reading source rows.
* @param target Adapter for writing rows to target.
* @param opts Migration options.
* @param sourceHasVector True if the source tier supports vector columns.
*/
export async function runMigrateTier(
source: MigrationSource,
target: MigrationTarget,
opts: MigrateTierOptions,
sourceHasVector = false,
): Promise<MigrateTierResult> {
const { dryRun, allowNonEmpty, batchSize, onProgress } = opts;
const tablesToMigrate = getMigrationOrder();
// Preflight: gather row counts from source.
onProgress('[migrate-tier] Gathering source row counts...');
const sourceCounts = new Map<string, number>();
for (const table of tablesToMigrate) {
const n = await source.count(table);
sourceCounts.set(table, n);
}
// Log preflight summary.
onProgress('[migrate-tier] Tables to migrate:');
for (const table of tablesToMigrate) {
const n = sourceCounts.get(table) ?? 0;
onProgress(` ${table}: ${n.toString()} rows`);
}
for (const skipped of SKIP_TABLES) {
onProgress(` ${skipped}: SKIPPED (ephemeral/environment-specific)`);
}
// Vector column notice.
if (!sourceHasVector) {
onProgress(
'[migrate-tier] NOTE: Source tier has no pgvector support. ' +
'insights.embedding will be NULL on all migrated rows.',
);
}
if (dryRun) {
onProgress('[migrate-tier] DRY RUN — no writes will be made.');
const tables: TableMigrationResult[] = tablesToMigrate.map((t) => ({
table: t,
rowsMigrated: 0,
skipped: false,
}));
for (const skipped of SKIP_TABLES) {
tables.push({ table: skipped, rowsMigrated: 0, skipped: true, skipReason: 'ephemeral' });
}
return { tables, totalRows: 0, dryRun: true };
}
// Check preconditions before writing.
await checkTargetPreconditions(target, allowNonEmpty, tablesToMigrate);
const results: TableMigrationResult[] = [];
let totalRows = 0;
for (const table of tablesToMigrate) {
const sourceCount = sourceCounts.get(table) ?? 0;
if (sourceCount === 0) {
onProgress(`[migrate-tier] ${table}: 0 rows — skipping.`);
results.push({ table, rowsMigrated: 0, skipped: false });
continue;
}
onProgress(`[migrate-tier] ${table}: migrating ${sourceCount.toString()} rows...`);
let offset = 0;
let tableTotal = 0;
let lastSuccessfulId: string | undefined;
try {
while (offset < sourceCount) {
const rows = await source.readTable(table, { limit: batchSize, offset });
if (rows.length === 0) break;
const normalised = rows.map((row) => normaliseSourceRow(table, row, sourceHasVector));
await target.upsertBatch(table, normalised);
lastSuccessfulId = rows[rows.length - 1]?.['id'] as string | undefined;
tableTotal += rows.length;
offset += rows.length;
onProgress(
`[migrate-tier] ${table}: ${tableTotal.toString()}/${sourceCount.toString()} rows written`,
);
}
} catch (err) {
const errMsg = err instanceof Error ? err.message : String(err);
throw new Error(
`[migrate-tier] Failed on table "${table}" after ${tableTotal.toString()} rows ` +
`(last id: ${lastSuccessfulId ?? 'none'}). Error: ${errMsg}\n` +
`Remediation: Re-run with --allow-non-empty to resume (upsert is idempotent).`,
);
}
results.push({ table, rowsMigrated: tableTotal, skipped: false });
totalRows += tableTotal;
onProgress(`[migrate-tier] ${table}: done (${tableTotal.toString()} rows).`);
}
// Add skipped table records.
for (const skipped of SKIP_TABLES) {
results.push({
table: skipped,
rowsMigrated: 0,
skipped: true,
skipReason: 'ephemeral or environment-specific — re-issue on target',
});
}
onProgress(`[migrate-tier] Complete. ${totalRows.toString()} total rows migrated.`);
return { tables: results, totalRows, dryRun: false };
}