Hacker News Reader

Comment by crazygringo | Hacker News Reader

I could easily see it measuring the forces and weight on the drum every 5 seconds (or even every 10 ms) during the whole wash, to be able to produce charts of vibration patterns, that engineers could use to correlate with failure. Remember -- when you're spinning at high speed to wring out the water, it's actually some pretty crazy strong forces.

Or other things measured every ~second, like stuff related to the motor, temperatues, humidity, etc. and other diagnostics.

Seems really easy to generate a megabyte if you consider time series. Even easier if it's in XML or JSON rather than a CSV.

View on HN

When I worked in a storage group years ago, the system that controls swipe card access for a building generated something like 3TB of Java exceptions a month.

Because of the criticality, it was on high tier reliable SAN storage, replicated to a second site. IIRC, storage was like $80 Gb/mo.

Love how you said exceptions, rather than just logs. I am unfortunately (painfully) aware of exactly what you mean. The storage costs were the cherry on the top. But seriously, $240k and nobody raised a stink?

It’s one of those things that cloud storage helps with.

Because it was on prem, the chargeback model was associated with the business unit and not super granular. It got lumped in with another business function because it should be a trivial workload.

I found it when I was doing estimates for a new platform and the app’s growth numbers didn’t add up! Even at $240k, it wasn’t an obvious outlier.

You really think they're going to measure all that, upload it, send it to some expensive engineer, have them try to physically model the error, and then... what?

That might make sense during development, but there's no way they do that in a consumer product. If a part breaks they're just going to send out a replacement or the repair guy is gonna get some third party part. Recording that much detail would just be noise.

Even if they had specific parts sensors (doubt it, for costs), they could just process that locally and send up an error code, not the whole log.

I find that all pretty hard to believe, but if anyone has evidence to the contrary, I'd be glad to be proven wrong. I had a LG washing machine bought new a few years ago promising all sorts of bells and whistles and app integrations. But it was super janky and cheaply made, the app integration was terrible, the on board memory would lose its configured settings, the entire LCD broke after a few weeks... it was not what I would consider well-engineered at all. If it was sending a megabyte a day I'd just assume it was yet another bug, not some forward thinking QA.

I'm not really sure what you're asking:

Is he sure they're sending < 1 MB a day? Yeah.

Is he sure it's plausible it's measuring time series data? Yeah.

Is it plausible they measure vibration patterns? Yeah.

It's not about like "oh we'll send an electrical engineer to fix your specific vibration pattern", it's "we can collect data from the field to make forward-looking decisions": ex. maybe we switch supplier mix for replacement motors, and a data scientist ends up finding 6 months later that something changed in June 2023 where serviced washers in New York report dramatically more intense vibrations and now we know to go talk to the supplier who gained mix.

It's error logs for non-tech. YMMV on individual team quality if they actually follow-up. Conceptually, Big Data(tm) is something CEOs have been hearing regularly since, what, 2014? So definitely plausible.

I would suspect they measure everything and use almost none of it. Like most web services. This was a common complaint and low hanging fruit optimization, that people were storing metrics that never got read. Just in case.

It's cute you think there is some intelligence to this design... it's just whatever some PM/exec dreamed up and some low-level engineer implemented based on requirements. The data is likely just sitting there collecting digital dust.

Well if you claim warranty, I'd expect them to want to have that data.

Maybe they also just sell it together with your advertising id, why not use washing patterns for deanonimyzation ...

I would imagine that they just log everything. Serial number, temperature, which cycle is used, time of day, how long it takes to fill the washer, how long it takes to drain the washer. Everything. Put all data in a great big database. When something needs to be fixed and is covered by the warranty, mark that the failed part is associated with that serial number.

Then do some sort of a regression to discover what logged parameters are associated with what failure modes/broken parts. If washers that take less time to fill up have higher than normal failure rates for some elbow joint, that probably means that high water pressure causes the elbow joint to fail. If a certain elbow join's failure rate is simply correlated with the number of cycles, that tells you something different. If a certain elbow join has a high failure rate that's not associated with anything, that probably just means it's a shitty part. But you learn something.

By logging everything and running a regression analysis, when you develop next year's model, you know where to improve. Now when you tell an expensive engineer, "This elbow join failed on 1000 units of revision F. Make it fail on 100 or less units of revision G." you can also give them a starting point to work with.

I'm a software guy. If I get 10 crash dumps, and you don't tell me anything, I don't necessarily know what to work with. If you give me those same 10 crash dumps and tell me that 9 of them had the language set to Arabic or Hebrew I know it's probably a BOM bug. Same thing.

Or you just sell the data to ad companies and let them figure out how to get value from it.

If they're not doing that they should, albeit finding a way to make the additional cost minor.

Collecting all that data for analysis would be incredibly valuable, especially considering the wealth of analysis tools today.

"Even easier if it's in XML or JSON rather than a CSV."

Yes indeed, but compression algorithms are not that new.

Sure, but it's also easy to imagine an engineer just forgetting to or not bothering because it wasn't in the spec.

Reply on news.ycombinator.com