I work for a company that has operated like this for 20 years. The system goes down sometimes, but we can fix it in less than an hour. At worst the users get a longer coffee break.
A single click in the software can often generate 500 SQL queries, so if you go from 0.05 ms to 1 ms latency you add half a second to clicks in the UI and that would piss our users off.
Definitely not saying this is the best way to operate at all times. But SQL has a huge problem with false dependencies between queries and API:s that make it very difficult to pipeline queries, so my experience has been that I/O-bound applications easily become extremely sensitive to latency.
Most businesses dont require that kind of uptime though. If i killed or servers for a couple of hours between 02:00 and 04:00 every night probably nobody would notice for at least a year if it wasn’t for the alerts we’d get.
A single click in the software can often generate 500 SQL queries, so if you go from 0.05 ms to 1 ms latency you add half a second to clicks
Those queries don’t all have to be executed sequentially though, do they? Usually if you have that many queries, at least some of them are completely independent of the others and thus can execute concurrently.
You don’t even need threading for that, just non-blocking IO and ideally an event loop.
The catch is that they all need to run in the same transaction to be unaffected by other things going on in the database and to make updates atomic. A single transaction means a single connection, and ODBC/JDBC has no way of multiplexing or pipelining queries over a single connection.
It’s probably theoretically possible to run some things in different transactions. But with all the different layers and complexity of the code (including third party components and ORMs like Hibernate), understanding all the failure modes and possible concurrency issues becomes intractable.
I work for a company that has operated like this for 20 years. The system goes down sometimes, but we can fix it in less than an hour. At worst the users get a longer coffee break.
A single click in the software can often generate 500 SQL queries, so if you go from 0.05 ms to 1 ms latency you add half a second to clicks in the UI and that would piss our users off.
Definitely not saying this is the best way to operate at all times. But SQL has a huge problem with false dependencies between queries and API:s that make it very difficult to pipeline queries, so my experience has been that I/O-bound applications easily become extremely sensitive to latency.
I’m going to guess quite a people here work on businesses where “sometimes breaks, but fixed in less than an hour” isn’t good enough for reliability.
Yeah if you need even 99.9% uptime, the most downtime you can accept in a year is eight hours.
Most businesses dont require that kind of uptime though. If i killed or servers for a couple of hours between 02:00 and 04:00 every night probably nobody would notice for at least a year if it wasn’t for the alerts we’d get.
Those queries don’t all have to be executed sequentially though, do they? Usually if you have that many queries, at least some of them are completely independent of the others and thus can execute concurrently.
You don’t even need threading for that, just non-blocking IO and ideally an event loop.
The catch is that they all need to run in the same transaction to be unaffected by other things going on in the database and to make updates atomic. A single transaction means a single connection, and ODBC/JDBC has no way of multiplexing or pipelining queries over a single connection.
It’s probably theoretically possible to run some things in different transactions. But with all the different layers and complexity of the code (including third party components and ORMs like Hibernate), understanding all the failure modes and possible concurrency issues becomes intractable.