Zero downtime

in theory many servers support hot deployment. In reality most support it poorly.

The only real way to have zero downtime is to never upgrade anything and hope that the Good Lord smiles upon you and gives you systems without any bugs, networks that are completely reliable, hardware that never fails, and powerstations that suffer no brownouts.

And even then Murphy is sure to cause some harddisks to run out of space once in a while.

A reasonable alternative is to use clustering and upgrade the nodes one at a time, not bringing the next one down until the upgraded one is back up and functioning.

That of course does limit what you can do during an ugprade (some upgrades may cause conflicts between nodes if the nodes run different versions of the software).

But, be honest: how important is zero downtime?

Most likely a few minutes every once in a while (especially outside business hours) is quite acceptable.

I've seen a graph outlining the cost increase as downtime percentage goes down.

With each 9 added to the uptime percentile (so from 99.9 to 99.99) the cost increase can be as much as a factor 10 or more, at what point do you decide that the added few minutes per year of uptime are no longer worth the cost to your business?

Can you live with 1 day downtime per year (spread over the year, so half an hour per month)?

Or maybe with 2.5 hours per day (a few minutes per month)?

Or do you think you can't live with more than 15 minutes downtime per year (a few seconds per month)?

Where do you draw the line? What can it cost?

And remember there are always external factors you can't control causing downtime. If an aircraft crashes onto your datacenter, it's going down.

You could build it as a nuclear bombshelter, but what about the power and datalines?

You could use geographically distributed datacenters, but those still need some means of synchronising data.

jwentinga at 2007-7-12 14:50:53 >