Re:Zero-downtime Postgres migrations – the hard parts
by Sinjo on Wednesday May 27, 2015 @07:57PM
(#9615225)
We can't count on a user (either in the browser or using the API) waiting around that long. Even some of our own services time out more aggressively than that.
Those situations can be smoothed over with retries, but we try pretty hard to avoid delays like this in the first place.
It's true the table wasn't available for 15 seconds, but current connections weren't dropped/lost, so a robust interface with the API should survive a 15 second delay.
Shouldn't a system like this be built to handle long delays?