Hacker News Reader

Comment by ojosilva | Hacker News Reader

Just to make it clear, the issue the OP is pointing out is with whoever wrote this inexcusable piece of code in the book, not MongoDB itself. Even if the authors later clarify there's the possibility of error, just publishing such misleading and grotesque solution for creating transactions in a non ACID database is very poor judgement.

MongoDB should not be used for an online ordering system, period. But if the programmer had no better alternative than Mongo, then please use Mongo's atomic operations [1] and nested documents to make sure nasty Bad Things don't happen.

[1] https://docs.mongodb.com/manual/tutorial/model-data-for-atom...

View on HN

In the case of an ecommerce site, where multiple products can go in a cart, I don't see any real right way to do things with MongoDB. Atomic for one row just doesn't help much, and you can't nest arbitrary cart product mixes.

As you say, bad idea in the first place.

I was part of a team that successfully did inventory in mongo. The cart makes zero difference if Mongo is only storing the inventory. It's fine to atomically change a single inventory row.

Operations that span multiple rows can be safely performed with a bus in front of it.

You can't judge the idea. I don't think you've quite grasped it. I agree the book example isn't great, but that's not the technology's fault.

I am assuming a system where available inventory is debited when the CC transaction completes.

How do you do that with single row only atomic transactions? And without charging $100 for what the customer thought was qty 3 $25 rakes and qty 1 $25 shovel...but is now something less than that due to concurrent purchases from other customers?

"How do you do that with single row only atomic transactions?"

In this scheme, you should theoretically be able to implement transactional protocols (two phase commit and friends, see, e.g., http://dl.acm.org/citation.cfm?id=3027893) that ensure you do not actually sell two people the same thing, but may return an error at buy time that the inventory disappeared while you tried to buy it.

You pay a sometimes-very-high cost in throughput.

Of course, talking about mongodb, it's probably broken in a myriad of ways no matter what you do :)

Nailed it, but I'm curious where you think the throughout suffered; we didn't encounter that problem.

It was a distributed system. The inventory was kept in Mongo with atomic debits and credits applied with a bus in front. SQL existed for the actual order, but inventory was managed in Mongo fine.

I'd like two of this product? Increment reserved in Mongo atomically, fire a bus message that inventory was reserved. I'd now like just one? Again atomically remove one from reserved and fire another message. Simplified of course.

Two users can't reserve the same inventory, reserving the inventory is atomic, and everyone else occurs during reconciliation. Typical distributed systems.

> It was a distributed system.

So you can't have exactly once semantics. When you say 'bus' I assume you mean a message broker which at best gives you at least once semantics. What happens when the message is delivered multiple times?

That's just a broker; all message handlers must be idempotent. This sometimes means storing message IDs or creating one-direction state change conduits that fail subsequent duplicate messages.

To answer that question more succinctly, nothing special happens when a message is delivered more than once.

At least, when we didn't screw up :)

Those were fun days.

I would assume you propagate a transaction ID or expected row version (and persist transaction histories and/or version numbers along with the inventory item row after mutation).

And what happens when that shovel gets stolen from the warehouse before it can be shipped to the customer?

You have to handle over-selling no matter what. The "source-of-truth" of you inventory system is whatever is actually, physically sitting in a warehouse somewhere. Not what your database says.

Yes, you have to handle overselling for other reasons, but I don't see why that's a good reason to not protect against it when you can.

I think Amazon calls it Apology Based Computing or something like that. You're better off at scale writing out a transaction log, then seeing if actions were successful and compensating otherwise.

> I was part of a team that successfully did inventory in mongo.

Yes, sure, let's go all get PHDs in electronics, networking protocols, IPC and database architecture, so we can write reliable transactions for our webshop/CRM. Or I don't know, maybe use mature system multiple decades in production.

Reply on news.ycombinator.com