Entries posted in March 2018

In my previous post I talked about how our child-care works here in wintery Finland, and suggested there might be a change in the near future.

So here is the predictable update; I've resigned from my job and I'm going to be taking over childcare/daycare. Ideally this will last indefinitely, but it is definitely going to continue until November. (Which is the earliest any child could be moved into public day-care if there problems.)

I've loved my job, twice, but even though it makes me happy (in a way that several other positions didn't) there is no comparison. Child-care makes me happier-still. Sure there are days when your child just wants to scream, refuse to eat, and nothing works. But on average everything is awesome.

It's a hard decision, a "brave" decision too apparently (which I read negatively!), but also an easy one to make.

It'll be hard. I'll have no free time from 7AM-5PM, except during nap-time (11AM-1PM, give or take). But it will be worth it.

And who knows, maybe I'll even get to rant at people who ask "Where's his mother?" I live for those moments. Truly.

Tags: bytemark, childcare, finland | 4 comments

I've been thinking about serverless-stuff recently, because I've been re-deploying a bunch of services and some of them could are almost microservices. One thing that a lot of my things have in common is that they're all simple HTTP-servers, presenting an API or end-point over HTTP. There is no state, no database, and no complex dependencies.

These should be prime candidates for serverless deployment, but at the same time I don't want to have to recode them for AWS Lamda, or any similar locked-down service. So docker is the obvious answer.

Let us pretend I have ten HTTP-based services, each of which each binds to port 8000. To make these available I could just setup a simple HTTP front-end:

• https://api.example.fi/

We'd need to route the request to the appropriate back-end, so we'd start to present URLs like:

• https://api.example.fi/steve/foo
• https://api.example.fi/steve/bar

Here any request which had the prefix steve/foo would be routed to a running instance of the docker container steve/foo. In short the name of the (first) path component performs the mapping to the back-end.

I wrote a quick hack, in golang, which would bind to port 80 and dynamically launch the appropriate containers, then proxy back and forth. I soon realized that this is a terrible idea though! The problem is a malicious client could start making requests for things like:

• https://api.example.fi/wordpress/wordpress
• https://api.example.fi/blah/blah

That would trigger my API-proxy to download the containers and spin them up. Allowing running arbitrary (albeit "sandboxed") code. So taking a step back, we want to use the path-component of an URL to decide where to route the traffic? Each container will bind to :8000 on its private (docker) IP? There's an obvious solution here: HAProxy.

So I started again, I wrote a trivial golang deamon which will react to docker events - containers starting and stopping - and generate a suitable haproxy configuration file, which can then be used to reload haproxy.

The end result is that if I launch a container named "foo" then requests to http://api.example.fi/foo will reach it. Success! The only downside to this approach is that you must manually launch your back-end docker containers - but if you do so they'll become immediately available.

I guess there is another advantage. Since you're launching the containers (manually) you can setup links, volumes, and what-not. Much more so than if your API layer span them up with zero per-container knowledge.

Tags: docker, golang, serverless | 1 comment

The past couple of days I've been reworking a few of my existing projects, and converting them from Perl into Golang.

Bytemark had a great alerting system for routing alerts to different enginners, via email, SMS, and chat-messages. The system is called mauvealert and is available here on github.

The system is built around the notion of alerts which have different states (such as "pending", "raised", or "acknowledged"). Each alert is submitted via a UDP packet getting sent to the server with a bunch of fields:

• Source IP of the submitter (this is implicit).
• A human-readable ID such as "heartbeat", "disk-space-/", "disk-space-/root", etc.
• A raise-field.
• More fields here ..

Each incoming submission is stored in a database, and events are considered unique based upon the source+ID pair, such that if you see a second submission from the same IP, with the same ID, then any existing details are updated. This update-on-receive behaviour is pretty crucial to the way things work, especially when coupled with the "raise"-field.

A raise field might have values such as:

• +5m
  • This alert will be raised in 5 minutes.
• now
  • This alert will be raised immediately.
• clear
  • This alert will be cleared immediately.

One simple way the system is used is to maintain heartbeat-alerts. Imagine a system sends the following message, every minute:

• id:heartbeat raise:+5m [source:1.2.3.4]
  • The first time this is received by the server it will be recorded in the database.
  • The next time this is received the existing event will be updated, and crucially the time to raise an alert will be bumped (i.e. it will become current-time + 5m).
  • The next time the update is received the raise-time will also be bumped
  • ..

At some point the submitting system crashes, and five minutes after the last submission the alert moves from "pending" to "raised" - which will make it visible in the web-based user-interface, and also notify an engineer.

With this system you could easily write trivial and stateless ad-hoc monitoring scripts like so which would raise/clear :

 curl https://example.com && send-alert --id http-example.com --raise clear --detail "site ok" || \
  send-alert  --id http-example.com --raise now --detail "site down"

In short mauvealert allows aggregation of events, and centralises how/when engineers are notified. There's the flexibility to look at events, and send them to different people at different times of the day, decide some are urgent and must trigger SMSs, and some are ignorable and just generate emails .

(In mauvealert this routing is done by having a configuration file containing ruby, this attempts to match events so you could do things like say "If the event-id contains "failed-disc" then notify a DC-person, or if the event was raised from $important-system then notify everybody.)

I thought the design was pretty cool, and wanted something similar for myself. My version, which I setup a couple of years ago, was based around HTTP+JSON, rather than UDP-messages, and written in perl:

• https://github.com/skx/purple

The advantage of using HTTP+JSON is that writing clients to submit events to the central system could easily and cheaply be done in multiple environments for multiple platforms. I didn't see the need for the efficiency of using binary UDP-based messages for submission, given that I have ~20 servers at the most.

Anyway the point of this blog post is that I've now rewritten my simplified personal-clone as a golang project, which makes deployment much simpler. Events are stored in an SQLite database and when raised they get sent to me via pushover:

• https://github.com/skx/purppura/

The main difference is that I don't allow you to route events to different people, or notify via different mechanisms. Every raised alert gets sent to me, and only me, regardless of time of day. (Albeit via an pluggable external process such that you could add your own local logic.)

I've written too much already, getting sidetracked by explaining how neat mauvealert and by extension purple was, but also I rewrote the Perl DNS-lookup service at https://dns-api.org/ in golang too:

• https://github.com/skx/dns-api-go

That had a couple of regressions which were soon reported and fixed by a kind contributor (lack of CORS headers, most obviously).

Tags: dns-api.org, golang, perl, purple, purppura, rewrite | 2 comments