Multi-document transactions are the right answer, and they are unremarkable on a replica set now:

1const session = await mongoose.startSession()
2try {
3  await session.withTransaction(async () => {
4    const [question] = await Question.create([doc], { session })
5    await TagQuestion.insertMany(links, { session })
6    await Tag.bulkWrite(ops, { session })
7  })
8} finally {
9  await session.endSession()
10}

1const session = await mongoose.startSession()
2try {
3  await session.withTransaction(async () => {
4    const [question] = await Question.create([doc], { session })
5    await TagQuestion.insertMany(links, { session })
6    await Tag.bulkWrite(ops, { session })
7  })
8} finally {
9  await session.endSession()
10}

withTransaction also retries on transient errors, which hand-rolled commit/abort blocks usually forget.

Two caveats: they require a replica set (Atlas gives you one; a bare local mongod does not), and they are more expensive than single-document writes — so use them where atomicity actually matters, not everywhere.

If you cannot use transactions, the fallback is to make the operation idempotent and reconcile — write the question first, then have a background job repair missing tag links. Strictly worse, but it is what you do on a sharded setup where the transaction cost is real.