> brodsky.dev_

Azure functions can look at blob storage and react to things.

But actually not really all that well.

Excerpt from the Documentation:

When you’re using a blob trigger on a Consumption plan, there can be up to a 10-minute delay in processing new blobs after a function app has gone idle. After the function app is running, blobs are processed immediately. To avoid this initial delay, consider one of the following options:

Use an App Service plan with Always On enabled.

Use another mechanism to trigger the blob processing, such as a queue message that contains the blob name. For an example, see Queue trigger with blob input binding.

Let’s deconstruct this a bit.

The important parts are the "Consumption Plan" vs "App Service", and how those relate to the Always On mode.

See, Azure Functions have two methods of operation (“plans”). The “Consumption” plan executes the function only when triggered. So if nothing is calling it, the function will go to sleep. A Function runs ephemerally and you need not think of its underlying resources whatsoever, aside from paying per invocation.

The App Service plan, on the other hand, launches a VM that will host your functions, and that VM remains running. You don’t need to directly manage it (nor can you), but you are being charged for all the minutes it’s humming away. Also, unlike the Consumption plan, you need to manage autoscaling yourself.

Only on the App Service plan you are given the option to enable “Always On”, which will prevent your function apps from going to sleep.

So in contrast to the probably familiar pattern of AWS Lambda being triggered by a change in S3 bucket, the Azure Blob storage doesn’t immediately trigger your function on change in blob storage, unless the function is already running. Otherwise, you are waiting for the scheduled wake-up window (feel free to correct me on Twitter if I am misunderstanding something). I personally find this behaviour to be super confusing, and inferior to what the rest of the cloud has come to to expect of the “serverless” patterns.

Good morning. Today we will take the terms “domains”, “fault”, and “update”, and make it sound more sophisticateder because competitive advantage.
- Azure marketing people, probably

I mean, it’s good they have thought of this. It’s even on the exam. But really, as the user of Azure, I don’t need to care about how they power their racks and in what order they are restarted. I care about stability of my VMs, but it’s ok to leave the mechanics of fault-tolerance to be a black box. For the most part, it would suffice for me to know that if I launch a group of 3 machines, I’ll have almost 3 machines running most of the time. I don’t have any control over this anyway, so those “domains” are trivia and implementation details.

That aside, Microsoft’s general aversion to visual presentation of data rears its ugly head here once again. They could have designed the UX around this as a nice grid, with current status of each slot in the fault/update domain, etc. Could’ve even put this next to each VM. But no. Everything must look like a spreadsheet.

The important takeaway of the entire feature: You should, for best availability vs cost effectiveness, try to horizontally scale your VMs in sets of 5: N % 5 == 0. That’s how many update domains exist. N < 5 - and you’re not utilizing the full fault-tolerance potential. 5 < N < 10 - and you are overprovisioning some of those update domains.