From: cross@spitfire.i.gajendra.net
In article <10f10gl$16kvh$5@dont-email.me>,
Lawrence D˙Oliveiro wrote:
>On Tue, 11 Nov 2025 18:56:53 -0500, Arne Vajhøj wrote:
>
>> HA is about whether the system can continue to serve users in case part
>> of a box or an entire box fail - 24x7 vs 16x5 is about architecture.
>
>High availability is measured in “nines” -- e.g. five nines, six nines ...
>even seven nines.
I don't normally reply to the troll, but, in this case multiple
factual misstatements deserve to be corrected.
>How do big enterprises (like Google) achieve that? By not using
>mainframes.
This has absolutely nothing to do with it. Google used COTS x86
gear because of cost, period. The software was then architected
to make this work well, and reliably.
Google achieves high availability because its internal systems
have been architected that way. But doing so is incredibly
expensive, in multiple dimensions, and the solutions are unique
to Google.
>They set up data centres full of off-the-shelf PC hardware --
This has not been true for two decades. Google designs and
manufactures its own computers for its datacenters. They are
nowhere close to COTS systems anymore.
>one article I remember from over a decade ago said that Google, at that
>time, had 460,000 servers.
Google has O(10^7) CPUs in O(10^6) computers in spread across
O(10^2) data centers, distributed globally. There are multiple
layers of redundancy and load balancing spreading traffic around
and routing around problems (which pop up regularly at that
scale). It also has automated monitoring, some automated
recovery, and and a small army of SREs and data center
technicians keeping everything running. It's not magic.
>All the hardware is obtained as cheaply as possible,
This has not been true for 15+ years. There was a time, early
in Google's life, when this was true, but those days are long
gone. Google has a highly developed, _highly_ skilled, internal
platforms team that designs and builds its own hardware at
nearly all levels of the stack. Very little is off the shelf
anymore, and none of it is "cheap".
>except one component:
>the power supply. They buy quality for that, for power-efficiency reasons.
>As for the rest, it doesn’t matter if a box falls over every minute, or a
>hard drive crashes every few minutes; they have higher-level redundancy
>and recovery procedures that can routinely recover from all those
>failures, without the users ever noticing.
This is true, but also nearly unique to the workloads Google
puts on its systems.
>No mainframe can match that.
Totally an apples and oranges comparison.
Google doesn't run workloads that look anything at all like what
traditionally runs on a mainframe.
- Dan C.
--- SoupGate-Win32 v1.05
* Origin: you cannot sedate... all the things you hate (1:229/2)
|
