@dforge-core/dforge-mcp 0.2.45 → 0.2.46-next

package/skills/dforge-1c-import/scripts/lib/delta.py

@@ -0,0 +1,741 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+"""
+Bring a loaded tenant up to date with what 1C changed since the load: the cutover's delta.
+
+    python tools/1c.py delta [module ...]            # dry run: what changed, per table, timed
+    python tools/1c.py delta [module ...] --apply    # write it
+    python tools/1c.py delta --status                # the watermarks
+    python tools/1c.py delta --baseline [module ...] # mark a tenant loaded before marks existed
+    python tools/1c.py delta --since 2026-09-17      # dry run as if loaded on that day
+
+What 1C gives to detect change was measured, not assumed (docs/design-cutover.md section 3.1):
+
+- `_Version` (rowversion) is on the OBJECT tables only - catalogs, documents, charts of
+  accounts and characteristic types. Tabular parts, register movements and information
+  registers have none. It is database-wide and monotonic, and none of the 472 tables
+  carrying it has an index on it - but a full scan of all of them for `_Version >= w` takes
+  half a second, so none is needed.
+- A deletion leaves no row. It is found by key: every object key 1C holds (2.6 M) is read
+  once per run into a temporary table, and a target row whose key is not there was deleted.
+- A tabular part changes only with its owner: 1C writes the object whole, so a changed line
+  bumps the owner's `_Version`. The lines of a changed owner are replaced.
+- A register movement changes only with its recorder, the same way: posting, reposting and
+  unposting write the document. The movements of a changed recorder are replaced - read
+  through the recorder index (`_RecorderTRef, _RecorderRRef`) every register carries.
+- What has neither - an independent information register, an aggregate the map reads
+  through a query - is read whole and compared row for row. They are small, bar two.
+
+So each loaded entity is one of four kinds, `object`, `lines`, `movements` and `whole`,
+decided from the column map and the schema dump, never named here.
+
+Every row goes through `migrate`'s own planning and conversions (`plan_entity`, `converted`,
+`reading`), into a staging table shaped like the target, and is compared there as text, so a
+row 1C touched without changing reads as unchanged. The apply runs in replica mode, as the
+load does, and ends the same way: `fill_missing` stands a placeholder in for a reference a
+deletion left dangling, `seedstate` reseeds a balance whose journal moved, and the
+`posted` markers are converged on the documents (that needs a token).
+
+The watermark is kept in the tenant, in `df1c.watermark`, one row per loaded entity: the
+`MIN_ACTIVE_ROWVERSION()` 1C reported before the rows were read. Below it every
+transaction had committed, so the next delta reads `_Version >= mark` and misses nothing.
+In the tenant rather than a file, because it is a fact about the tenant's rows - written in
+the transaction that wrote them, so neither can survive without the other, and a second
+tenant loaded from the same infobase keeps its own. `migrate run` writes it too.
+
+A delta assumes the target holds only what 1C gave it: a row keyed by nothing 1C knows is
+taken for a deletion. While 1C is the system of record, nothing is entered in the tenant.
+"""
+from __future__ import print_function
+
+import argparse
+import datetime
+import re
+import sys
+import time
+
+import migrate
+import paths
+
+SCHEMA = "df1c"
+KEYS_IN_LIST = 500
+# Charts of calculation types (`_CKinds`) are objects too - `_IDRRef` and `_Version` like a
+# catalog. Leaving them out read every row of the four the standard configuration loads as
+# deleted, since their keys never reached the set of live ones.
+OBJECT = re.compile(r"^_(Document|Reference|Acc|Chrc|CKinds|Node|BPr|Task)\d+$")
+OWNER = re.compile(r"^(_(?:Document|Reference|Acc|Chrc|CKinds|Node)\d+)_IDRRef$")
+
+
+# --------------------------------------------------------------------------- watermarks
+
+_READY = set()
+
+
+def ensure(pg):
+	if id(pg) in _READY:
+		return
+	cur = pg.cursor()
+	cur.execute("CREATE SCHEMA IF NOT EXISTS %s" % SCHEMA)
+	cur.execute("""CREATE TABLE IF NOT EXISTS %s.watermark (
+		module     text NOT NULL,
+		entity     text NOT NULL,
+		source     text NOT NULL,
+		version    bytea NOT NULL,
+		updated_at timestamptz NOT NULL DEFAULT now(),
+		PRIMARY KEY (module, entity))""" % SCHEMA)
+	# Entities whose rows a delta moved and whose module has not been settled since - its
+	# placeholders, balances and markers. Written in the same transaction as the rows, so a
+	# run that dies between the two leaves the next one something to finish.
+	cur.execute("""CREATE TABLE IF NOT EXISTS %s.unsettled (
+		module     text NOT NULL,
+		entity     text NOT NULL,
+		since      timestamptz NOT NULL DEFAULT now(),
+		PRIMARY KEY (module, entity))""" % SCHEMA)
+	_READY.add(id(pg))
+
+
+def unsettle(pg, module, job):
+	pg.cursor().execute("INSERT INTO %s.unsettled (module, entity) VALUES (%%s, %%s)"
+	                    " ON CONFLICT DO NOTHING" % SCHEMA, (module, job.entity))
+
+
+def unsettled(pg, module):
+	ensure(pg)
+	cur = pg.cursor()
+	cur.execute("SELECT entity FROM %s.unsettled WHERE module = %%s" % SCHEMA, (module,))
+	return set(r[0] for r in cur.fetchall())
+
+
+def horizon(mssql):
+	"""Everything below this `_Version` is committed; anything written later is at or above."""
+	mssql.execute("SELECT MIN_ACTIVE_ROWVERSION()")
+	return bytes(mssql.fetchone()[0])
+
+
+def record(pg, module, job, version, lower_only=False):
+	"""Set an entity's mark, inside the caller's transaction."""
+	ensure(pg)
+	cur = pg.cursor()
+	cur.execute("""INSERT INTO {s}.watermark AS w (module, entity, source, version)
+	               VALUES (%s, %s, %s, %s)
+	               ON CONFLICT (module, entity) DO UPDATE
+	               SET version = {v}, source = EXCLUDED.source, updated_at = now()"""
+	            .format(s=SCHEMA, v="LEAST(w.version, EXCLUDED.version)" if lower_only
+	                    else "EXCLUDED.version"),
+	            (module, job.entity, job.source, as_binary(version)))
+
+
+def marks(pg):
+	ensure(pg)
+	cur = pg.cursor()
+	cur.execute("SELECT module, entity, version, updated_at FROM %s.watermark" % SCHEMA)
+	return dict(((m, e), (bytes(v), at)) for m, e, v, at in cur.fetchall())
+
+
+def as_binary(value):
+	import psycopg2
+	return psycopg2.Binary(value)
+
+
+def hexed(raw):
+	return "0x" + bytes(raw).hex()
+
+
+# --------------------------------------------------------------------------- classification
+
+def classify(job, source, mapped):
+	"""(kind, detail) for one planned entity. See the module docstring for the four kinds.
+
+	An entity whose query joins in a column from ANOTHER table - the latest status of a
+	shipping order, read off an information register that other documents write - is read
+	whole: its row changes when that table does, and the owner's `_Version` does not move.
+	"""
+	if mapped.get("joined"):
+		return "whole", None
+	columns = source[job.source]["columns"]
+	versioned = (columns.get("_Version") or {}).get("type") == "timestamp"
+	exposes = not job.query or re.search(r"\b\w+\.\*", job.query)
+	keys = [c for c, s in job.pairs if s == "_IDRRef"]
+	if versioned and keys and exposes:
+		return "object", keys[0]
+	for column, src in job.pairs:
+		match = OWNER.match(src)
+		if match and (source.get(match.group(1), {}).get("columns", {})
+		              .get("_Version", {}).get("type") == "timestamp"):
+			return "lines", (column, src, match.group(1))
+	recorder = [c for c, s in job.pairs if s == "_RecorderRRef"]
+	if recorder:
+		return "movements", (recorder[0], "_RecorderTRef" in columns)
+	return "whole", None
+
+
+def table_number(name):
+	return int(re.search(r"(\d+)$", name).group(1))
+
+
+# --------------------------------------------------------------------------- what 1C holds
+
+class Infobase(object):
+	"""Every object key 1C holds and its `_Version`, read once per run.
+
+	2.7 M keys across 472 tables; the read is a few seconds and is the whole of the change
+	detection for objects, lines and movements.
+	"""
+
+	def __init__(self, mssql, source, since):
+		self.versions = {}          # table -> [(raw key, raw version)] at or after `since`
+		self.now = horizon(mssql)
+		started = time.time()
+		self.keys = 0
+		self.alive = []
+		for name, info in sorted(source.items()):
+			if not OBJECT.match(name):
+				continue
+			if (info["columns"].get("_Version") or {}).get("type") != "timestamp":
+				continue
+			if "_IDRRef" not in info["columns"]:
+				continue
+			# Two reads rather than one: the key alone comes off the narrowest index,
+			# while `_Version` lives only in the clustered one. Measured on 2.7 M keys:
+			# 90 s for (key, version) over every row, a few seconds this way.
+			mssql.execute("SELECT _IDRRef, _Version FROM [%s] WHERE _Version >= %s"
+			              % (name, hexed(since)))
+			self.versions[name] = [(bytes(k), bytes(v)) for k, v in mssql.fetchall()]
+			mssql.execute("SELECT _IDRRef FROM [%s]" % name)
+			while True:
+				batch = mssql.fetchmany(50000)
+				if not batch:
+					break
+				self.keys += len(batch)
+				self.alive.extend(bytes(r[0]) for r in batch)
+		self.seconds = time.time() - started
+
+	def changed(self, table, mark):
+		return [k for k, v in self.versions.get(table, ()) if v >= mark]
+
+	def changed_documents(self, mark):
+		"""{table number: [raw key]} over every document table: the recorders that moved."""
+		out = {}
+		for name, rows in self.versions.items():
+			if not name.startswith("_Document"):
+				continue
+			keys = [k for k, v in rows if v >= mark]
+			if keys:
+				out[table_number(name)] = keys
+		return out
+
+	def load_alive(self, pg):
+		"""The key set into a temporary table, for the anti-joins that find deletions."""
+		started = time.time()
+		cur = pg.cursor()
+		cur.execute("DROP TABLE IF EXISTS pg_temp.df1c_alive")
+		cur.execute("CREATE TEMP TABLE df1c_alive (id uuid)")
+		for at in range(0, len(self.alive), 200000):
+			migrate.copy_rows(pg, "pg_temp", "df1c_alive", ["id"],
+			                  [[migrate.to_uuid(k)] for k in self.alive[at:at + 200000]])
+		cur.execute("CREATE UNIQUE INDEX ON pg_temp.df1c_alive (id)")
+		cur.execute("ANALYZE pg_temp.df1c_alive")
+		return time.time() - started
+
+
+# --------------------------------------------------------------------------- staging
+
+def q(name):
+	return '"%s"' % name.replace('"', '""')
+
+
+def stage(pg, mssql, job, columns, restricts):
+	"""Read the rows `restricts` select from 1C through the load's own conversions into a
+	temporary table shaped like the target. `restricts` is a list of alias -> condition
+	functions, each one query (a long key list is split), or [None] for the whole source."""
+	cur = pg.cursor()
+	cur.execute("DROP TABLE IF EXISTS pg_temp.df1c_stage")
+	cur.execute('CREATE TEMP TABLE df1c_stage AS SELECT %s FROM "%s"."%s" WITH NO DATA'
+	            % (", ".join(q(c) for c in columns), job.schema, job.table))
+	select = ", ".join("[%s]" % p[1] for p in job.pairs)
+	staged = 0
+	for restrict in restricts:
+		mssql.execute("SELECT %s FROM %s" % (select, migrate.reading(job, restrict)))
+		while True:
+			batch = mssql.fetchmany(migrate.CHUNK)
+			if not batch:
+				break
+			rows = migrate.converted(job, batch)
+			migrate.copy_rows(pg, "pg_temp", "df1c_stage", columns, rows)
+			staged += len(rows)
+	cur.execute("ANALYZE pg_temp.df1c_stage")
+	return staged
+
+
+def in_lists(keys, column, alias):
+	"""`alias.column IN (0x.., ...)`, one condition per slice of the key list."""
+	out = []
+	for at in range(0, len(keys), KEYS_IN_LIST):
+		listed = ", ".join(hexed(k) for k in keys[at:at + KEYS_IN_LIST])
+		out.append(lambda a, listed=listed: "%s.%s IN (%s)" % (a, column, listed))
+	return out
+
+
+def recorder_lists(changed, typed):
+	"""Restrictions selecting the movements of the changed recorders, through the recorder
+	index: by (TRef, RRef) where the register is written by several kinds, by RRef where it
+	is written by one."""
+	if not typed:
+		keys = [k for group in changed.values() for k in group]
+		return in_lists(keys, "_RecorderRRef", None)
+	out = []
+	for number, keys in sorted(changed.items()):
+		tref = "0x%08x" % number
+		for cond in in_lists(keys, "_RecorderRRef", None):
+			out.append(lambda a, tref=tref, cond=cond:
+			           "%s._RecorderTRef = %s AND %s" % (a, tref, cond(a)))
+	return out
+
+
+def texts(columns, alias):
+	return ", ".join("%s.%s::text" % (alias, q(c)) for c in columns)
+
+
+# --------------------------------------------------------------------------- one entity
+
+class Result(object):
+	def __init__(self, module, job, kind):
+		self.module, self.job, self.kind = module, job, kind
+		self.changed = self.new = self.deleted = 0
+		self.owners = 0            # lines/movements: owners or recorders replaced
+		self.staged = 0
+		self.read = self.compare = self.write = 0.0
+		self.note = ""
+
+	@property
+	def touched(self):
+		return bool(self.changed or self.new or self.deleted)
+
+
+def has_missing(pg, job):
+	return any(c[0] == "is_missing" for c in migrate.table_columns(pg, job.schema, job.table))
+
+
+def run_entity(pg, mssql, lookup, base, module, job, kind, detail, mark, apply, now):
+	res = Result(module, job, kind)
+	job = migrate.bind_enums(job, lookup)
+	columns = [p[0] for p in job.pairs] + ([job.derived] if job.derived else [])
+	keys = sorted(migrate.primary_key(pg, job.schema, job.table))
+	missing = has_missing(pg, job)
+	real = " AND NOT t.is_missing" if missing else ""
+	target = '"%s"."%s"' % (job.schema, job.table)
+	cur = pg.cursor()
+	started = time.time()
+
+	if kind == "object":
+		key_column = detail
+		res.staged = stage(pg, mssql, job, columns,
+		                   [lambda a: "%s._Version >= %s" % (a, hexed(mark))])
+		res.read = time.time() - started
+		started = time.time()
+		on = " AND ".join("t.%s = s.%s" % (q(k), q(k)) for k in keys)
+		cur.execute("SELECT count(*) FILTER (WHERE t.%s IS NULL%s),"
+		            " count(*) FILTER (WHERE t.%s IS NOT NULL%s AND ROW(%s) IS DISTINCT FROM ROW(%s))"
+		            " FROM pg_temp.df1c_stage s LEFT JOIN %s t ON %s"
+		            % (q(keys[0]), " OR t.is_missing" if missing else "",
+		               q(keys[0]), " AND NOT t.is_missing" if missing else "",
+		               texts(columns, "s"), texts(columns, "t"), target, on))
+		res.new, res.changed = cur.fetchone()
+		gone = ("FROM %s t WHERE NOT EXISTS (SELECT 1 FROM pg_temp.df1c_alive k"
+		        " WHERE k.id = t.%s)%s" % (target, q(key_column), real))
+		cur.execute("SELECT count(*) " + gone)
+		res.deleted = cur.fetchone()[0]
+		res.compare = time.time() - started
+		if apply and res.touched:
+			started = time.time()
+			begin(cur)
+			updates = [c for c in columns if c not in keys]
+			sets = ", ".join("%s = EXCLUDED.%s" % (q(c), q(c)) for c in updates)
+			if missing:
+				sets += (", " if sets else "") + "is_missing = false"
+			into = columns + (["is_missing"] if missing else [])
+			cur.execute("INSERT INTO %s AS t (%s) SELECT %s%s FROM pg_temp.df1c_stage"
+			            " ON CONFLICT (%s) DO %s"
+			            % (target, ", ".join(q(c) for c in into),
+			               ", ".join(q(c) for c in columns), ", false" if missing else "",
+			               ", ".join(q(k) for k in keys),
+			               ("UPDATE SET %s WHERE ROW(%s%s) IS DISTINCT FROM ROW(%s%s)"
+			                % (sets, texts(columns, "t"), ", t.is_missing" if missing else "",
+			                   ", ".join("EXCLUDED.%s::text" % q(c) for c in columns),
+			                   ", false" if missing else "")) if sets else "NOTHING"))
+			cur.execute("DELETE " + gone)
+			unsettle(pg, module, job)
+			record(pg, module, job, now)
+			cur.execute("COMMIT")
+			res.write = time.time() - started
+		elif apply:
+			begin(cur)
+			record(pg, module, job, now)
+			cur.execute("COMMIT")
+		return res
+
+	if kind in ("lines", "movements"):
+		if kind == "lines":
+			owner, owner_src, owner_table = detail
+			changed = base.changed(owner_table, mark)
+			restricts = in_lists(changed, owner_src, None)
+			changed_ids = [migrate.to_uuid(k) for k in changed]
+		else:
+			owner, typed = detail
+			by_type = base.changed_documents(mark)
+			restricts = recorder_lists(by_type, typed)
+			changed_ids = [migrate.to_uuid(k) for group in by_type.values() for k in group]
+		# Owners the target holds that 1C no longer has: their rows go too.
+		cur.execute("DROP TABLE IF EXISTS pg_temp.df1c_owners")
+		cur.execute("CREATE TEMP TABLE df1c_owners (id uuid)")
+		started_gone = time.time()
+		# The anti-join first and the DISTINCT over what survives it - next to nothing -
+		# rather than a DISTINCT over the whole table: 75 s against 24 M postings otherwise.
+		# More memory for this one statement only. Left on the session it outlived its
+		# purpose and a later parallel hash join asked for more shared memory than the
+		# container's /dev/shm has ("could not resize shared memory segment").
+		cur.execute("SET work_mem = '256MB'")
+		cur.execute("INSERT INTO pg_temp.df1c_owners SELECT DISTINCT g.id FROM"
+		            " (SELECT t.%s AS id FROM %s t WHERE t.%s IS NOT NULL AND NOT EXISTS"
+		            " (SELECT 1 FROM pg_temp.df1c_alive k WHERE k.id = t.%s) OFFSET 0) g"
+		            % (q(owner), target, q(owner), q(owner)))
+		gone_owners = cur.rowcount
+		cur.execute("RESET work_mem")
+		gone_seconds = time.time() - started_gone
+		if changed_ids:
+			migrate.copy_rows(pg, "pg_temp", "df1c_owners", ["id"], [[i] for i in changed_ids])
+		cur.execute("CREATE INDEX ON pg_temp.df1c_owners (id)")
+		cur.execute("ANALYZE pg_temp.df1c_owners")
+		res.owners = len(changed_ids) + gone_owners
+		res.staged = stage(pg, mssql, job, columns, restricts) if restricts else 0
+		if not restricts:
+			cur.execute("DROP TABLE IF EXISTS pg_temp.df1c_stage")
+			cur.execute('CREATE TEMP TABLE df1c_stage AS SELECT %s FROM %s WITH NO DATA'
+			            % (", ".join(q(c) for c in columns), target))
+		res.read = time.time() - started - gone_seconds
+		started = time.time()
+		# What the target holds for those owners, taken once: the table may have no index on
+		# the owner column (the ledger has none on its recorder), and both directions of the
+		# comparison read it.
+		cur.execute("DROP TABLE IF EXISTS pg_temp.df1c_old")
+		cur.execute("CREATE TEMP TABLE df1c_old AS SELECT %s FROM %s t"
+		            " WHERE t.%s IN (SELECT id FROM pg_temp.df1c_owners)"
+		            % (", ".join("t.%s" % q(c) for c in columns), target, q(owner)))
+		subset = "SELECT %s FROM pg_temp.df1c_old t" % texts(columns, "t")
+		staged = "SELECT %s FROM pg_temp.df1c_stage s" % texts(columns, "s")
+		cur.execute("SELECT (SELECT count(*) FROM (%s EXCEPT ALL %s) a),"
+		            " (SELECT count(*) FROM (%s EXCEPT ALL %s) b)"
+		            % (staged, subset, subset, staged))
+		res.new, res.deleted = cur.fetchone()
+		res.compare = time.time() - started + gone_seconds
+		if apply and res.touched:
+			started = time.time()
+			begin(cur)
+			cur.execute("DELETE FROM %s t WHERE t.%s IN (SELECT id FROM pg_temp.df1c_owners)"
+			            % (target, q(owner)))
+			cur.execute("INSERT INTO %s (%s) SELECT %s FROM pg_temp.df1c_stage"
+			            % (target, ", ".join(q(c) for c in columns),
+			               ", ".join(q(c) for c in columns)))
+			unsettle(pg, module, job)
+			record(pg, module, job, now)
+			cur.execute("COMMIT")
+			res.write = time.time() - started
+		elif apply:
+			begin(cur)
+			record(pg, module, job, now)
+			cur.execute("COMMIT")
+		return res
+
+	# whole: read it all, compare by key.
+	res.staged = stage(pg, mssql, job, columns, [None])
+	res.read = time.time() - started
+	started = time.time()
+	on = " AND ".join("t.%s = s.%s" % (q(k), q(k)) for k in keys)
+	cur.execute("SELECT count(*) FILTER (WHERE t.%s IS NULL%s),"
+	            " count(*) FILTER (WHERE t.%s IS NOT NULL%s AND ROW(%s) IS DISTINCT FROM ROW(%s))"
+	            " FROM pg_temp.df1c_stage s LEFT JOIN %s t ON %s"
+	            % (q(keys[0]), " OR t.is_missing" if missing else "",
+	               q(keys[0]), " AND NOT t.is_missing" if missing else "",
+	               texts(columns, "s"), texts(columns, "t"), target, on))
+	res.new, res.changed = cur.fetchone()
+	gone = ("FROM %s t WHERE NOT EXISTS (SELECT 1 FROM pg_temp.df1c_stage s WHERE %s)%s"
+	        % (target, on, real))
+	cur.execute("SELECT count(*) " + gone)
+	res.deleted = cur.fetchone()[0]
+	res.compare = time.time() - started
+	if apply and res.touched:
+		started = time.time()
+		begin(cur)
+		updates = [c for c in columns if c not in keys]
+		sets = ", ".join("%s = EXCLUDED.%s" % (q(c), q(c)) for c in updates)
+		if missing:
+			sets += (", " if sets else "") + "is_missing = false"
+		into = columns + (["is_missing"] if missing else [])
+		cur.execute("DELETE " + gone)
+		cur.execute("INSERT INTO %s AS t (%s) SELECT %s%s FROM pg_temp.df1c_stage"
+		            " ON CONFLICT (%s) DO %s"
+		            % (target, ", ".join(q(c) for c in into),
+		               ", ".join(q(c) for c in columns), ", false" if missing else "",
+		               ", ".join(q(k) for k in keys),
+		               ("UPDATE SET %s WHERE ROW(%s) IS DISTINCT FROM ROW(%s)"
+		                % (sets, texts(columns, "t"),
+		                   ", ".join("EXCLUDED.%s::text" % q(c) for c in columns)))
+		               if sets else "NOTHING"))
+		unsettle(pg, module, job)
+		record(pg, module, job, now)
+		cur.execute("COMMIT")
+		res.write = time.time() - started
+	elif apply:
+		begin(cur)
+		record(pg, module, job, now)
+		cur.execute("COMMIT")
+	return res
+
+
+def begin(cur):
+	"""One transaction per entity, in replica mode as the load runs: the generated lock
+	and posted-document guards would refuse exactly these rows, and so would a foreign key
+	to a row another module brings in later in the same delta."""
+	cur.execute("BEGIN")
+	cur.execute("SET LOCAL session_replication_role = 'replica'")
+
+
+# --------------------------------------------------------------------------- after a module
+
+def settle(pg, module, results, apply, mark_documents):
+	"""What a module needs once its rows moved: placeholders, balances, markers."""
+	import posted
+	import seedstate
+	touched = set(r.job.entity for r in results if r.touched)
+	if apply:
+		touched |= unsettled(pg, module)
+	schemas = sorted(set(r.job.schema for r in results))
+	notes = []
+	settled = True
+	if apply and touched:
+		cur = pg.cursor()
+		cur.execute("BEGIN")
+		cur.execute("SET LOCAL session_replication_role = 'replica'")
+		filled = migrate.fill_missing(pg, schemas, verbose=False)
+		cur.execute("COMMIT")
+		if filled:
+			notes.append("%d placeholder(s)" % sum(f[4] for f in filled))
+
+	cur = pg.cursor()
+	for reg in seedstate.registers(cur, set([module])):
+		if reg.journal_entity not in touched:
+			continue
+		if apply:
+			pg.autocommit = False
+			rows, written, _ = seedstate.seed(pg, reg)
+			pg.autocommit = True
+			off = seedstate.drift(pg.cursor(), reg)
+			settled = settled and not off
+			notes.append("reseeded %s.%s: %d tuple(s), check %s"
+			             % (reg.module, reg.code, written,
+			                "clean" if not off else "OFF BY %d" % off))
+		else:
+			notes.append("would reseed %s.%s" % (reg.module, reg.code))
+
+	# A totals cache is not maintained by a bulk write any more than by the load: the
+	# platform folds movements in only at posting. `totals` refills it from the journal with
+	# the platform's own SQL, transcribed, and checks it the platform's way.
+	cur.execute("""SELECT r.register_cd, je.entity_cd FROM "dForge".register r
+	               JOIN "dForge".module m ON m.module_id = r.module_id
+	               JOIN "dForge".entity je ON je.entity_id = r.journal_entity_id
+	               WHERE m.module_cd = %s AND r.totals IS NOT NULL""", (module,))
+	stale = set(code for code, journal in cur.fetchall() if journal in touched)
+	if stale:
+		import totals
+		for plan in totals.plans(pg.cursor(), set([module])):
+			if plan.code not in stale:
+				continue
+			if not apply:
+				notes.append("would rebuild the totals cache of %s.%s" % (module, plan.code))
+				continue
+			totals.rebuild(pg, plan)
+			_, off = totals.check(pg, plan, 5)
+			settled = settled and not off
+			notes.append("rebuilt the totals cache of %s.%s, check %s"
+			             % (module, plan.code, "clean" if not off else "OFF BY %d" % off))
+
+	work, _, _ = posted.survey(pg, module, quiet=True)
+	work = [w for w in work if w[1].split(".", 1)[1] in touched]
+	if work:
+		if apply and mark_documents:
+			done, ok = posted.converge(pg, work, mark_documents)
+			settled = settled and ok
+			notes.append("markers: %s%s" % (", ".join("%s %d" % kv for kv in done.items()) or
+			                                "nothing to do", "" if ok else " (FAILED)"))
+		elif apply:
+			settled = False
+			notes.append("markers NOT converged (no token): run `posted --mark %s`, or this"
+			             " again with a token" % module)
+		else:
+			notes.append("would converge markers on %d document table(s)" % len(work))
+	if apply and touched and settled:
+		pg.cursor().execute("DELETE FROM %s.unsettled WHERE module = %%s" % SCHEMA, (module,))
+	return notes
+
+
+# --------------------------------------------------------------------------- command
+
+def since_date(mssql, source, day):
+	"""A watermark for estimation: the lowest `_Version` of any document dated on or after
+	`day`. Everything 1C wrote from about then on is at or above it."""
+	shifted = "%04d-%s" % (int(day[:4]) + migrate.YEAR_OFFSET, day[5:])
+	lowest = None
+	for name, info in source.items():
+		if not name.startswith("_Document") or not OBJECT.match(name):
+			continue
+		if "_Date_Time" not in info["columns"] or \
+				(info["columns"].get("_Version") or {}).get("type") != "timestamp":
+			continue
+		mssql.execute("SELECT MIN(_Version) FROM [%s] WHERE _Date_Time >= '%s'" % (name, shifted))
+		found = mssql.fetchone()[0]
+		if found is not None and (lowest is None or bytes(found) < lowest):
+			lowest = bytes(found)
+	if lowest is None:
+		raise SystemExit("delta: no document dated on or after %s" % day)
+	return lowest
+
+
+def modules_in_order(mapping):
+	try:
+		import rebuild
+		order = [m for m in rebuild.install_order() if m in mapping]
+	except SystemExit:
+		order = []
+	return order + [m for m in mapping if m not in order]
+
+
+def main():
+	migrate.load_env()
+	ap = argparse.ArgumentParser(description=__doc__.strip().splitlines()[0])
+	ap.add_argument("module", nargs="*", help="limit to these modules (default: every one)")
+	ap.add_argument("--entity", action="append", help="limit to this entity (repeatable)")
+	ap.add_argument("--apply", action="store_true", help="write the delta (default: dry run)")
+	ap.add_argument("--dry-run", action="store_true", help="report only (the default)")
+	ap.add_argument("--since", help="override every watermark: YYYY-MM-DD (estimation) or 0x...")
+	ap.add_argument("--baseline", action="store_true",
+	                help="set the mark of every entity that has none to 1C's horizon now")
+	ap.add_argument("--status", action="store_true", help="list the watermarks")
+	ap.add_argument("--no-mark", action="store_true",
+	                help="--apply without converging the posted markers")
+	args = ap.parse_args()
+	if args.apply and args.dry_run:
+		raise SystemExit("delta: --apply and --dry-run together")
+
+	mapping, source = migrate.column_map(), migrate.source_schema()
+	pg = migrate.pg_connect()
+	pg.autocommit = True          # no read transaction held open on the metadata tables
+	modules = [m for m in modules_in_order(mapping) if not args.module or m in args.module]
+	unknown = set(args.module) - set(mapping)
+	if unknown:
+		raise SystemExit("delta: not in the column map: %s" % ", ".join(sorted(unknown)))
+	known = marks(pg)
+
+	if args.status:
+		for (module, entity), (version, at) in sorted(known.items()):
+			if module in modules:
+				print("%-12s %-40s %s  %s" % (module, entity, hexed(version), at))
+		return 0
+
+	db = migrate.mssql_connect()
+	mssql, lookup = db.cursor(), db.cursor()
+	plan = []
+	for module in modules:
+		for job in migrate.jobs_for(pg, module, mapping, source, args.entity):
+			if job.blocking:
+				continue
+			plan.append((module, job))
+
+	if args.baseline:
+		now = horizon(mssql)
+		set_ = 0
+		cur = pg.cursor()
+		cur.execute("BEGIN")
+		for module, job in plan:
+			if (module, job.entity) not in known:
+				record(pg, module, job, now)
+				set_ += 1
+		cur.execute("COMMIT")
+		print("baseline %s set on %d entit(ies); %d already had one."
+		      % (hexed(now), set_, len(plan) - set_))
+		print("Only right if the tenant holds what 1C holds NOW - straight after a full load"
+		      " during which 1C did not change.")
+		return 0
+
+	override = None
+	if args.since:
+		override = (bytes.fromhex(args.since[2:]) if args.since.startswith("0x")
+		            else since_date(mssql, source, args.since))
+	unmarked = [(m, j.entity) for m, j in plan if (m, j.entity) not in known]
+	if unmarked and override is None:
+		raise SystemExit("delta: %d entit(ies) have no watermark (e.g. %s.%s) - load them with"
+		                 " `migrate run`, or `delta --baseline` if the tenant already matches 1C"
+		                 % (len(unmarked), unmarked[0][0], unmarked[0][1]))
+
+	def mark_of(module, job):
+		return override if override is not None else known[(module, job.entity)][0]
+
+	lowest = min(mark_of(m, j) for m, j in plan)
+	t0 = time.time()
+	base = Infobase(mssql, source, lowest)
+	alive_seconds = base.load_alive(pg)
+	print("1C horizon %s, reading from %s%s" % (hexed(base.now), hexed(lowest),
+	                                             " (--since %s)" % args.since if args.since else ""))
+	print("object keys: %d read in %.1fs, %.1fs into the tenant"
+	      % (base.keys, base.seconds, alive_seconds))
+	print("%s\n" % ("APPLYING" if args.apply else "dry run - nothing written; --apply writes it"))
+
+	jwt = None
+	if args.apply and not args.no_mark:
+		try:
+			import posting
+			jwt = posting.token()
+		except SystemExit as missing:
+			print("no token (%s) - markers will be left for `posted --mark`" % missing)
+
+	header = "%-11s %-34s %-9s %8s %8s %8s %8s %7s %7s %7s" % (
+		"module", "entity", "kind", "owners", "changed", "new", "deleted",
+		"read s", "cmp s", "write s")
+	print(header)
+	print("-" * len(header))
+	totals = [0, 0, 0, 0, 0.0]
+	all_notes = []
+	for module in modules:
+		results = []
+		for m, job in plan:
+			if m != module:
+				continue
+			kind, detail = classify(job, source, mapping[module][job.entity])
+			res = run_entity(pg, mssql, lookup, base, module, job, kind, detail,
+			                 mark_of(module, job), args.apply, base.now)
+			results.append(res)
+			totals[0] += res.changed
+			totals[1] += res.new
+			totals[2] += res.deleted
+			totals[3] += res.owners
+			totals[4] += res.read + res.compare + res.write
+			if res.touched or res.owners or res.read + res.compare > 2 or args.entity:
+				print("%-11s %-34s %-9s %8s %8d %8d %8d %7.1f %7.1f %7.1f"
+				      % (module, job.entity[:34], kind,
+				         res.owners if kind in ("lines", "movements") else "",
+				         res.changed, res.new, res.deleted, res.read, res.compare, res.write))
+				sys.stdout.flush()
+		for note in settle(pg, module, results, args.apply, jwt if args.apply else None):
+			all_notes.append("%s: %s" % (module, note))
+	print("-" * len(header))
+	print("%d entities  changed %d  new %d  deleted %d  (owners/recorders replaced %d)"
+	      % (len(plan), totals[0], totals[1], totals[2], totals[3]))
+	print("entity work %.0fs, whole run %.0fs" % (totals[4], time.time() - t0))
+	for note in all_notes:
+		print("  " + note)
+	print("(rows of an entity with nothing changed and under 2s are not listed;"
+	      " lines/movements count rows replaced as new/deleted)")
+	return 0
+
+
+if __name__ == "__main__":
+	sys.exit(main() or 0)