On a latest (9.0.0-alpha0-d6d27ec8) sharded cluster, opening a change stream and issuing:

{aggregate: 1, pipeline: [{$changeStream: …}], cursor: {}}

returns its first (empty) batch after a delay that differs by server version:

• 8.3 sharded: ~5–13ms, every time (6/6).
• latest sharded: ~1.0s (driver evergreen tools sets periodicNoopIntervalSecs=1 🤷), most of the time (~5/6; the rest ~10ms).

The slow case still succeeds, returning an empty firstBatch and a postBatchResumeToken after ~1.0s. It's not an error from the server, but is a delay.

Repro:

  for (let i = 0; i < 6; i++) {
    const t0 = Date.now();
    const r = db.getSiblingDB("admin").runCommand({
      aggregate: 1,
      pipeline: [{ $changeStream: { allChangesForCluster: true } }],
      cursor: {}
    });
    print(`open #${i}: ${Date.now() - t0}ms  ok=${r.ok}  firstBatch=${r.cursor.firstBatch.length}`);
    sleep(250);
  }