[andrewxdiamond]

Ironically the _lack_ of dogfooding GCP products at google is often quoted as one of the reasons AWS beat GCP to defining the Cloud market. Amazon builds AWS on AWS as much as possible, Google has only somewhat recently pushed for this

[GuB-42]

What I understood is that AWS is more than dogfooding. It is something Amazon first built for themselves, to give more independence to individual teams. And as they noticed it worked well, they realized that they could turn it into a product.

For what I understand as an outsider, Google is much more monolithic, having a platform where each team can do their things independently is not really their culture, so if they build one, it is only for their customers, because they don't work like this internally. Whereas for Amazon, an AWS customer is not that different from one of their own teams.

[ouddv]

That’s mostly a marketing myth on the AWS side. As recently as three or four years ago there were _new_ initiatives being built in the legacy “corp” fabric; and even today Amazon has internal tooling that makes use of Native AWS quite different than it is for external customers; particularly around authn/authz.

And that doesn’t even mention the comic “Moving to AWS” platform that technically consumed AWS resources, but was a wholly different developer experience to native.

[nateglims]

Now building on AWS inside is heavily emphasized, but just a few years ago most services were built with internal systems that are very different. Some solutions (multi account/cellular architecture for example) seemed to come from dog fooding heavily, but supporting services (like account SSO for handling many accounts) are still very different from the publicly available equivalents.

[oceanplexian]

As someone who worked at AWS it’s ironic how hard they dog food cellular architecture but when it comes to customers, all the offerings and docs are terrible, with the only information in obscure Re:Invent talks or blog posts.

I now work for a large customer and you would be shocked at the household names that basically put all their infrastructure in a single Account and Region. Or they have multi region but it’s basically an afterthought and wouldn’t serve any purpose in a disaster.

[randomcarbloke]

Catfooding

[nextos]

I think Gmail was great initially because of dogfooding. Right now, the incentives are different, and it's more about releasing new stuff. And we can see how that worked with the Google Chat saga.

Lots of other Google products suffer from similar issues because of an apparent lack of dogfooding. I bought a Pixel phone not so long ago and I had to install all updates, one by one, to bring it to the latest Android version. It took several days.

[derefr]

I can see why they do it, though. There are a bunch of foundational Google infra technologies that are great for building an IaaS on top of, but which can't themselves be offered as IaaS services for whatever reason.

Let's use Google's Colossus (their datacenter-scale virtual filesystem) as an example. Due to the underlying architecture of Colossus, GCP can turn around and give you:

• GCE shared read-only zonal PDs

• near-instantaneous snapshots for GCE and BigTable

• async and guaranteed-durable logging (for GCE and otherwise) and Queues (as Pub/Sub and otherwise)

• zero-migration autoclassed GCS Objects, and no per-operation slowdown on GCS Buckets as bucket size increases

• BigQuery being entirely serverless (vs e.g. Redshift needing to operate on a provisioned-storage model)

But Google can't just sell you "Colossus as a service" — because Colossus doesn't have a "multitenant with usage-cost-based backpressure to disincentivize misuse" architecture; and you can't add that without destroying the per-operation computational-complexity guarantees that make Colossus what it is. Colossus only works in a basically-trusted environment. (A non-trust-requiring version of Colossus would look like Apple's FoundationDB.)

(And yeah, you could in theory have a "little Colossus" unique to your deployment... but that'd be rather useless, since the datacenter scale of Colossus is rather what makes many of its QoS guarantees possible. Though I suppose it could make sense if you could fund entire GCP datacenters for your own use, ala AWS GovCloud.)

[antonvs]

Probably more importantly, doesn't the Amazon store system use AWS? Google has nothing comparable to use for that purpose.

[hmottestad]

There is search, Adsense, gmail, google docs and Gemini. Do they at least train Gemini on GPUs on GCP?

[bobthepanda]

maps is another big one.

[pie420]

one of the craziest comments i've read on HN. google does a lot of internet things these days, idk if you've been out of the loop for a while

[antonvs]

I didn't express it well.

Google's consumer-facing systems all tend to be very focused. Things like search, maps, gmail etc. are not the same kind of system as Amazon's store.

While these systems do presumably give Google something to exercise their cloud systems on, the sense I have (as a longtime user of both GCP and AWS) is that it doesn't give them a realistic sense of what other companies, that don't just sell advertising and consumer data via focused products, do. Amazon's store is more representative of typical businesses in that sense.

Basically, it seems to me that Google Cloud has continually learned lessons the hard way about what customers need, rather than getting that information from its own internal usage.