diff --git a/app/server/lib/ActiveDoc.ts b/app/server/lib/ActiveDoc.ts
index c87baca4..56f4ec83 100644
--- a/app/server/lib/ActiveDoc.ts
+++ b/app/server/lib/ActiveDoc.ts
@@ -27,7 +27,7 @@ import {toTableDataAction} from 'app/common/DocActions';
 import {DocData} from 'app/common/DocData';
 import {DocSnapshots} from 'app/common/DocSnapshot';
 import {EncActionBundleFromHub} from 'app/common/EncActionBundle';
-import {byteString} from 'app/common/gutil';
+import {byteString, countIf} from 'app/common/gutil';
 import {InactivityTimer} from 'app/common/InactivityTimer';
 import * as marshal from 'app/common/marshal';
 import {Peer} from 'app/common/sharing';
@@ -1125,13 +1125,44 @@ export class ActiveDoc extends EventEmitter {
    * collaborators.
    */
   private async _migrate(docSession: OptDocSession): Promise<void> {
-    // TODO fetchAllTables() creates more memory pressure than usual since full data is present in
-    // memory at once both in node and in Python. This is liable to cause crashes. This doesn't
-    // even skip onDemand tables. We should try to migrate using metadata only. Data migrations
-    // would need to be done table-by-table, and differently still for on-demand tables.
-    const allTables = await this.docStorage.fetchAllTables();
-    const docActions: DocAction[] = await this._dataEngine.pyCall('create_migrations', allTables);
-    this.logInfo(docSession, "_migrate: applying %d migration actions", docActions.length);
+    const tableNames = await this.docStorage.getAllTableNames();
+
+    // Fetch only metadata tables first, and try to migrate with only those.
+    const tableData: {[key: string]: Buffer|null} = {};
+    for (const tableName of tableNames) {
+      if (tableName.startsWith('_grist_')) {
+        tableData[tableName] = await this.docStorage.fetchTable(tableName);
+      }
+    }
+
+    let docActions: DocAction[];
+    try {
+      // The last argument tells create_migrations() that only metadata is included.
+      docActions = await this._dataEngine.pyCall('create_migrations', tableData, true);
+    } catch (e) {
+      if (!/need all tables/.test(e.message)) {
+        throw e;
+      }
+      // If the migration failed because it needs all tables (i.e. involves changes to data), then
+      // fetch them all. TODO: This is used for some older migrations, and is relied on by tests.
+      // If a new migration needs this flag, more work is needed. The current approach creates
+      // more memory pressure than usual since full data is present in memory at once both in node
+      // and in Python; and it doesn't skip onDemand tables. This is liable to cause crashes.
+      this.logWarn(docSession, "_migrate: retrying with all tables");
+      for (const tableName of tableNames) {
+        if (!tableData[tableName] && !tableName.startsWith('_gristsys_')) {
+          tableData[tableName] = await this.docStorage.fetchTable(tableName);
+        }
+      }
+      docActions = await this._dataEngine.pyCall('create_migrations', tableData);
+    }
+
+    const processedTables = Object.keys(tableData);
+    const numSchema = countIf(processedTables, t => t.startsWith("_grist_"));
+    const numUser = countIf(processedTables, t => !t.startsWith("_grist_"));
+    this.logInfo(docSession, "_migrate: applying %d migration actions (processed %s schema, %s user tables)",
+      docActions.length, numSchema, numUser);
+
     docActions.forEach((action, i) => this.logInfo(docSession, "_migrate: docAction %s: %s", i, shortDesc(action)));
     await this.docStorage.execTransaction(() => this.docStorage.applyStoredActions(docActions));
   }
diff --git a/app/server/lib/DocStorage.ts b/app/server/lib/DocStorage.ts
index 505f04d6..fe0f5de1 100644
--- a/app/server/lib/DocStorage.ts
+++ b/app/server/lib/DocStorage.ts
@@ -714,22 +714,11 @@ export class DocStorage implements ISQLiteDB {
   }
 
   /**
-   * Fetches all tables from the database, and returns a dictionary mapping table names to encoded
-   * TableData objects (marshalled dicts mapping column ids to arrays of values).
+   * Fetches and returns the names of all tables in the database (including _gristsys_ tables).
    */
-  public fetchAllTables(): Promise<{[key: string]: Buffer|null}> {
-    const tables: {[key: string]: Buffer|null} = {};
-    return bluebird.Promise.each(
-      this.all("SELECT name FROM sqlite_master WHERE type='table'"),
-      (row: ResultRow) => {
-        if (!row.name.startsWith('_gristsys_')) {
-          return this.fetchTable(row.name)
-            .then(data => {
-              tables[row.name] = data;
-            });
-        }
-      })
-      .return(tables);
+  public async getAllTableNames(): Promise<string[]> {
+    const rows = await this.all("SELECT name FROM sqlite_master WHERE type='table'");
+    return rows.map(row => row.name);
   }
 
   /**
diff --git a/sandbox/grist/main.py b/sandbox/grist/main.py
index 75b53896..16742c18 100644
--- a/sandbox/grist/main.py
+++ b/sandbox/grist/main.py
@@ -97,9 +97,9 @@ def main():
     return eng.load_table(table_data_from_db(table_name, table_data))
 
   @export
-  def create_migrations(all_tables):
+  def create_migrations(all_tables, metadata_only=False):
     doc_actions = migrations.create_migrations(
-      {t: table_data_from_db(t, data) for t, data in all_tables.iteritems()})
+      {t: table_data_from_db(t, data) for t, data in all_tables.iteritems()}, metadata_only)
     return map(actions.get_action_repr, doc_actions)
 
   @export
diff --git a/sandbox/grist/migrations.py b/sandbox/grist/migrations.py
index 4c984a78..664e836f 100644
--- a/sandbox/grist/migrations.py
+++ b/sandbox/grist/migrations.py
@@ -28,12 +28,18 @@ all_migrations = {}
 def noop_migration(_all_tables):
   return []
 
+# Each migration function includes a .need_all_tables attribute. See migration() decorator.
+noop_migration.need_all_tables = False
 
-def create_migrations(all_tables):
+
+def create_migrations(all_tables, metadata_only=False):
   """
   Creates and returns a list of DocActions needed to bring this document to
-  schema.SCHEMA_VERSION. It requires as input all of the documents tables.
-    all_tables: all tables as a dictionary mapping table name to TableData.
+  schema.SCHEMA_VERSION.
+    all_tables: all tables or just the metadata tables (those named with _grist_ prefix) as a
+      dictionary mapping table name to TableData.
+    metadata_only: should be set if only metadata tables are passed in. If ALL tables are
+      required to process migrations, this method will raise a "need all tables..." exception.
   """
   try:
     doc_version = all_tables['_grist_DocInfo'].columns["schemaVersion"][0]
@@ -57,6 +63,7 @@ def create_migrations(all_tables):
   new_schema = {a.table_id: a for a in schema.schema_create_actions()}
   for table_id, data in sorted(all_tables.iteritems()):
     # User tables should already be in tdset; the rest must be metadata tables.
+    # (If metadata_only is true, there is simply nothing to skip here.)
     if table_id not in tdset.all_tables:
       new_col_info = {}
       if table_id in new_schema:
@@ -71,6 +78,9 @@ def create_migrations(all_tables):
 
   migration_actions = []
   for version in xrange(doc_version + 1, schema.SCHEMA_VERSION + 1):
+    migration_func = all_migrations.get(version, noop_migration)
+    if migration_func.need_all_tables and metadata_only:
+      raise Exception("need all tables for migration to %s" % version)
     migration_actions.extend(all_migrations.get(version, noop_migration)(tdset))
 
   # Note that if we are downgrading versions (i.e. doc_version is higher), then the following is
@@ -86,12 +96,19 @@ def get_last_migration_version():
   """
   return max(all_migrations.iterkeys())
 
-def migration(schema_version):
+def migration(schema_version, need_all_tables=False):
   """
   Decorator for migrations that associates the decorated migration function with the given
-  schema_version. This decorate function will be run to migrate forward to schema_version.
+  schema_version. This decorated function will be run to migrate forward to schema_version.
+
+  Migrations are first attempted with only metadata tables, but if any required migration function
+  is marked with need_all_tables=True, then the migration will be retried with all tables.
+
+  NOTE: new migrations should NOT set need_all_tables=True; it would require more work to process
+  very large documents safely (incuding those containing on-demand tables).
   """
   def add_migration(migration_func):
+    migration_func.need_all_tables = need_all_tables
     all_migrations[schema_version] = migration_func
     return migration_func
   return add_migration
@@ -653,7 +670,9 @@ def migration16(tdset):
 
   return tdset.apply_doc_actions(doc_actions)
 
-@migration(schema_version=17)
+# This is actually the only migration that requires all tables because it modifies user data
+# (specifically, any columns of the deprecated "Image" type).
+@migration(schema_version=17, need_all_tables=True)
 def migration17(tdset):
   """
   There is no longer an "Image" type for columns, as "Attachments" now serves as a