Hacker News Reader

Comment by RestlessMind | Hacker News Reader

> One of the features contained a bug which would cause the GFE to restart; this bug had not been detected in either of testing and initial rollout.

I wish they elaborated more on what type of bug was that, which was not caught by testing / initial rollout. Either tests must be poorly written or the bug must be very subtle.

On an unrelated note: kudos to Google for publishing this postmortem and hope that this becomes an industrywide practice. I also wish they publish (a belated) one about Google+ and their throng of messaging apps over the years.

View on HN

With all due respect, it's not a postmortem, it's an advert. It doesn't really say anything other than "We had a problem, we fixed it.". There are virtually no technical details in there other than "something would restart spontaneously, which shifted the load somewhere else". Maybe i'm a bit jaded by cloudflare and aws writeups, but this really isn't anything special or worthwhile reading.

In this case, they might be hard-pressed to go into details, since it was related to an unreleased feature:

« These features had been introduced into the second layer GFE code base but not yet put into service. One of the features contained a bug which would cause the GFE to restart; this bug had not been detected in either of testing and initial rollout. »

...but yeah, they definitely could have done better.

Another reason postmortems can be vague is that, for each detail you add, you might need to add even more background details. Then it becomes a recursive problem.

In this case, the "configuration change" could be a feature flag in the L2 GFEs, something in the L1 GFEs that changed L1->L2 requests in minor ways, or maybe something else entirely (since it's partially security-related: a dynamic LOAS handshake change to use different cyphers? So many possibilities). At the end of the day, though, it's still a specific permutation of all possible features and flags that hadn't been vetted before. Given how large Google's monorepo is, it's not impossible for one of your many dependencies, even indirect ones, which might be configured by another team entirely, to have subtle time bombs that only trigger well after you have built and deployed the code.

Having been on the other side, I know that, for every detail added, a bunch more questions come up.

Is there a write up that explains what the technologies are that you mentioned in your post? LOAS, L1 GFE, L2 GFE, ... never heard of those

L1 and L2 GFE are the Level 1 and Level 2 Google Front Ends which the post mortem talks about. I'm not familiar with LOAS.

LOAS is how Stubby (gRPC) connections on the internal network are secured, looks a bit like mTLS and is partially being open sourced as ALTS: https://cloud.google.com/security/encryption-in-transit/appl...

GFE (Google Front End) is what you connect to when you access any Google services through your browser. Think nginx or even ELB. It's a load balancing reverse proxy, as well as WAF. It's mentioned here https://cloud.google.com/security/infrastructure/design/ and probably in the SRE book. This report might be the first time Google mentions in public that there can be two levels of GFEs, but I remember at least one service using such a setup many, many years ago.

Very little detail. I'm sure they're worried about disclosing something about their stack that would open them up to more vulnerability which is understandable, but not explaining why they didn't see it in their testing... That's something they could disclose without any threats.

It's easier to disclose details of a physical infrastructure outage or a bug in someone else's code/product than a bug in your own code.

Given my recent experiences with Google, I have to say that I'm impressed that they supplied even that much detail.

That’s an interesting long term strategy to keep the bar so low that the slightest step is going above.

You've just described most people's career plans.

I donno, looks comprehensive enough for me. What were you looking for, exactly? Source code? Packet traces? Pew-pew maps?

A useful PM include a summary, impact analysis, root cause analysis, and a comprehensive (and realistic) set of measures that will prevent recurrence.

After reading what they provided, I have a reasonable understanding of what went wrong (sufficient for me to plan my own safeguards if necessary), a useful measure of the team's response and remediation capabilities, and enough information to judge my comfort with their preventative measures.

Anything else is just cleverly-disguised marketing.

It's pretty bare-bones next to a postmortem like https://www.epicgames.com/fortnite/en-US/news/postmortem-of-...

I feel that's way too much info.

Actually "We had a problem, we fixed it, we made sure it won't happen again."

That last bit would be more marketing except that it's true - I've used Google Appengine for years, and the rare outages are always unique issues.

I'd love to see comparative numbers, but my impression is that this focus on improvement has lead to Google's uptime being a lot better than AWS or Azure.

Totally agree with you on this one. While they keep absolute technical details to themselves and not on the page, at least give a link or something where interested people can read what went on.

Also, if I give similar root cause in my environment, I'd be laughed off. We need to absolutely provide what went wrong, what was the immediate fix and what's the permanent fix and is it done (or does it require downtime / restart).

Google internally will totally have a document with stack traces and code snippets.

They have just chosen to only post a brief summary.

To your point, all postmortems are advertisements. This is just an unconvincing "me too" ad. Amazon and CloudFlare's quality and detailed write-ups give technical people more confidence in supporting products, as well as more interest in joining the company.

All postmortems are ads, but not all ads are effective.

This is a self congratulatory lazy... ad.

I agree. The AWS S3 outage postmortem had a much more specific description of the technical problem (even though it amounted to bad parameters being passed to a script).

To be fair, nowhere on the linked page does it say that this is a postmortem. The submitter called it that.

It's the perfect doc to show my leadership team, who want to know what happened, but aren't very technical.

> On an unrelated note: kudos to Google for publishing this postmortem and hope that this becomes an industrywide practice.

I think most already do this. I have seen AWS also publish detailed postmortems for outages like this. Ex: https://aws.amazon.com/message/41926/

I started tracking them a while back, didn't stick with it for very long.

https://github.com/macintux/Service-postmortems

This is super high level summary not a postmortem, in fact it doesn't say anything about the bug and why there was no test for it, what repair items are done, etc..

Amazon and Microsoft's post mortems are much more to the point, one can actually learn from them to not make the same or similar mistake.

I will note that the word "postmortem" appears nowhere on the page, which is dedicated to status updates about the incident in question.

It reads to me like the bug required a particular feature to be enabled, and the feature wasn't enable when the software was rolling out, then the feature was globally knife-switched to the "on" position without an orderly staged rollout.

I actually think it looks pretty bad that one of the action items in this report is to make a feature dashboard for GFEs. They've been saying they will do that for years, and the team that operates GFEs is considered the most elite of all SRE teams. Most famous outages of Google products have been caused by bogus configurations pushed to GFEs or network devices in front of them.

It's sort of like that extra disk in your raid 1 setup at home.

Compared to a single SSD the performance improvement doesn't really show (for desktop loads)...

However when things come to a rare (but somewhat inevitable) and screeching halt and that mirror has one of the copies shatter beyond recognition... that's when the doubled price proves that it was worth it.

Such a dashboard would invariably also add load and complexity (both failure points) to the system, but outwardly most users would be unaware of their existence.

As it says in the note, the bug was in a feature that was latent but not yet being used. Then a configuration change started hitting the feature.

With highly redundant systems such as this, you generally need multiple layers of things going wrong all at once to notice an issue. This was the case here as well.

  Either tests must be poorly written or
  the bug must be very subtle.

The easiest way to make a web service work in testing but fail in production is a problem with the settings that are necessarily different between environments.

For example, you must have different database credentials between test and production, and you must limit who can read the production credentials. If the production credentials are malformed, a service that worked in test will fail in production.

And the same applies to your SSL certificates, your settings for enabled/disabled features, your flashy markers that stop people mistaking production for test....

Sometimes bugs are triggered by very specific data or queries (aka "queries of death"). GFEs have many tests, so I'd go for the subtle bug.

I suppose every detected "query of death" should be included in the next version of the service's test suite.

I bet there is already a regression test for this one.

> ... this bug had not been detected in either of testing and initial rollout.

It makes me so happy[0] that the big smart corps have the same problems that us plebs have.

I tell people that testing can never precisely duplicate the production environment but do they believe me?

Also, this is an argument for feature switches vs staging environments.

[0] But not in a schadenfreude sense.

Another possibility is that the change was not tested at all or very minimally tested -- and they would not want to admit to that publicly.

Reply on news.ycombinator.com