I’ve been having trouble with insomnia for several years. It came in waves, a couple weeks it was better, a couple weeks worse. I’ve been using a sleep tracking app to keep track of how bad I slept. I’d look at it and analyze everything and feel sorry for myself.

Say it was like 3 years and say I spent on average 2 hours a night not being able to sleep. That’d be over 2000 hours lost, and that’s probably a conservative estimate.

I always knew one had to get up regularly, but haven’t been doing it since probably 2012. I’d try to limit blue-light in the evening, and did all kinds of things which are tangential to good sleep. Anything but regular sleep schedule!

About two weeks ago I decided enough was enough, deleted the sleep tracking app, set up a “dumb” alarm for 7:07 every day of the week, and reduced total time in bed to something uncomfortably short (around six and half hours). For the past week I’ve been sleeping great, all thanks to these two things. I still have trouble getting up at 7:07, but that’s a short moment of discomfort compared to hours of insomnia.

There’s a wonderful page describing how to sleep better. I didn’t even need to do all that much to cure my terrible insomnia – just the basic basics helped almost immediately. I wonder, how else can I easily drastically improve my quality of life?

Elm is a purely functional programming language for creating web frontend. It is strongly typed, largely influenced by Haskell – without some of the things that make Haskell difficult for newcomers, and with much better error messages.

I’ve recently written two small (~2k lines total) side projects in Elm. Writing code in Elm is an absolute joy:

• It’s easy to get started and Elm is a simple language.
• Very strong typing – seems somewhat stronger than in Haskell. Eg List.head is of type List a -> Maybe a and won’t crash on you in runtime – if the list is empty you get Nothing. If it compiles it usually works.
• Elm has amazing error messages. They don’t just tell you what’s wrong, they try to explain why that could be and suggest how you could try to fix the error. I have never before encountered errors this helpful.
• There’s Ellie – a wonderful online environment to play around and share snippets.
• The Elm Slack is the most helpful and responsive community I’ve ever participated in.
• Refactoring is a breeze – it compiles it works. I’ve never had a refactoring mess up things that worked before.
• Yes Elm compiles to JavaScript, and I’ve never had to dive into the generated code.

That said, there are some potential drawbacks:

• No higher kinded polymorphism: to map over a List, you need to use List.map, to map over Set, use Set.map. This is slightly verbose and sucks for the library developers, but isn’t a big problem for the end user.
• Elm-land is an autocracy. Evan has been a very enlightened ruler, so this could also be seen as a benefit.
• The development of Elm itself is intentionally slow: they try to get it right rather than get it out quick. Some things one would like in the standard library are external libraries. The repository of Elm packages is rather comprehensive.

Set contents of /etc/systemd/resolved.conf to:

[Resolve]
DNS=45.90.28.239 45.90.30.239
Domains=~.

Apparently spaces are separators. Those are NextDNS servers in the example btw. NextDNS has not sponsored this post.

Then run: $ sudo service systemd-resolved restart

This was one of about two hundred million things that took me too long to figure out while setting up Ubuntu 20.04.

And in case you want to resolve single-label lookups, systemd-resolved don’t do that.

• apt install dnsmasq resolvconf
• Set /etc/dnsmasq.conf config, eg values from my.nextdns.io
• Disable systemd-resolved
• Set dns=dnsmasq in /etc/NetworkManager/NetworkManager.conf
• Set nameserver 127.0.0.1 in /etc/resolvconf/resolv.conf.d/base
• Restart all teh things

I don’t start my showers cold, but always end them so. I start the shower warm, then switch to hot, then at the end as cold as possible for about half a minute. In the beginning it was mildly uncomfortable, but I’ve become accustomed to it quickly. Showering in different places, I’m often surprised at the range of different temperatures of the coldest water available.

A cold shower really wakes you up. It closes your pores, which is good for the skin, and prevents heat loss. After a cold shower, you’ll always feel great about yourself.

Interestingly, ending with cold showers is perhaps the only positive habit I’ve built recently. I usually find forming habits challenging. This one didn’t require much work at all – usually I’m good at finding excuses, the easiest one being “I don’t have time for this”, which doesn’t apply here. Yes, I do have half a minute (and so do you, you’re reading this after all).

This is gonna be rambling personal experience. You have been warned.

A year ago I took a break from my work at Inviqa to learn Scala and functional programming. This has been a great decision, a most successful endeavour, and an amazing learning experience.

I’ve been stuck with PHP for a long time. Mind, at Inviqa we were doing as good PHP as one can. I learned lots about software engineering from a multitude of conference speakers and industry leaders. This was further helped by a very active learning & development department and a generous conference/learning budget. Still, it was PHP. The language sucks.

Learning Scala as a “better Java” kind of language was easy. Case classes, options, lots of easy wins. Learning functional programming… well, I’ve been reading the awesome Functional Programming in Scala book, and I’m still reading it one year later. Despite its hands-on approach and lots of hand-holding, it’s not easy for me to digest. Functional programming is hard.

I used to really like Python. I wrote about 50 lines of Python recently, and just couldn’t believe how difficult it was. One needs to explicitly write return to return values. When just trying things out, you don’t get quick feedback from the compiler. Either write tests or run your code – the former seemed like overdoing it for such a short program, the latter was tedious. Working with the collections was a pain.

Let’s have a look at an example. What we want to do: split a string on commas, then strip whitespace from each element.

In Python:
map(lambda i: i.strip(), re.split(‘,’, string))

In Scala:
string.split(“,”).map(_.trim)

The relative conciseness of Scala is nice, sure, but the Scala code basically follows the instructions: take a string, split on a comma, trim whitespace. Python is completely cryptic in comparison.

The main drawback of Scala is not the language but the ecosystem. SBT (Scala Build Tool) is powerful but often times incomprehensible. The compilation can be slow at times. Scala’s Play Framework is not nearly as polished as PHP’s Symfony. Replacing any piece of Symfony is just a dependency injection away. Replacing a component of Play is bordering on impossible, everything depends on everything and is mostly hardcoded. I wanted to do a really simple one-line change in Play Framework routing but had to give up.

Inviqa didn’t have as much interest in Scala as I hoped, so regrettably, I stopped working for them. Now I’m part of KwiqJobs (soon to be Swarms Technologies), an amazing startup. We turn people’s waiting time into a new resource for companies. Our tech team works remotely and we meet about once a month for a week. It’s lots of fun! Also we’re hiring! :)

You probably don’t want to use a ListMap. It has an interface of a Map, but is “powered” (that’s too good a word) by a List. Thus pretty much any Map operations you call on it will take ages. Consider using a LinkedHashMap instead – it will preserve both the element order and your sanity.

That’s me trying to blog more.

I’ve heard the opinion that indexes on low cardinality columns don’t work well. I set out to disprove that.

First, let’s create a table with a boolean column “active” and fill it with dummy data. Please note it includes a (default, B-Tree) index on the “active” column.

> show create table users\G
*************************** 1. row ***************************
Table: users
Create Table: CREATE TABLE `users` (
`id` int(11) NOT NULL AUTO_INCREMENT,
`name` varchar(50) DEFAULT NULL,
`email` varchar(100) DEFAULT NULL,
`active` tinyint(1) DEFAULT NULL,
PRIMARY KEY (`id`),
KEY `active` (`active`)
) ENGINE=InnoDB DEFAULT CHARSET=latin1
1 row in set (0.00 sec)

> DELIMITER ;;
> CREATE PROCEDURE insert_random(IN amount INT, IN percent_active INT)
    -> BEGIN
    ->   DECLARE counter INT DEFAULT 0;
    ->   myloop: LOOP
    ->     if (counter = amount) THEN
    ->       LEAVE myloop;
    ->     end if;
    ->     INSERT INTO users SET
    ->       name = UUID(),
    ->       email = UUID(),
    ->       active = IF(FLOOR(RAND() * 100) < percent_active, 1, 0);
    ->     SET counter = counter + 1;
    ->   END LOOP;
    -> END ;;
Query OK, 0 rows affected (0.00 sec)

> DELIMITER ;

> call insert_random(1000000, 10);
Query OK, 1 row affected (36.60 sec)

> select count(*) from users;
+----------+
| count(*) |
+----------+
|  1000000 |
+----------+
1 row in set (0.18 sec)

We’ve inserted a million users into our table! That took some time. Please note the 0.18 seconds to count them all.

Now, let’s see how fast we can count the active users:

> select count(*) from users where active = true;
+----------+
| count(*) |
+----------+
|    99887 |
+----------+
1 row in set (0.04 sec)

Cool, pretty fast. Can you explain that?

> explain select count(*) from users where active = true\G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: users
   partitions: NULL
         type: ref
possible_keys: active
          key: active
      key_len: 2
          ref: const
         rows: 208462
     filtered: 100.00
        Extra: Using index
1 row in set, 1 warning (0.00 sec)

Now let’s drop the index and run the test again:

> alter table users drop key active;
Query OK, 0 rows affected (0.05 sec)
Records: 0  Duplicates: 0  Warnings: 0

> select count(*) from users where active = true;
+----------+
| count(*) |
+----------+
|    99887 |
+----------+
1 row in set (0.24 sec)

Wow, a lot slower… I’ve ran the selects a couple of times with consistent results to verify no caching was influencing the results. Let’s see the explain:

> explain select count(*) from users where active = true\G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: users
   partitions: NULL
         type: ALL
possible_keys: NULL
          key: NULL
      key_len: NULL
          ref: NULL
         rows: 912236
     filtered: 10.00
        Extra: Using where
1 row in set, 1 warning (0.00 sec)

As you hopefully know: “Using index” good, “Using where” bad.

TL;DR A boolean column consisting of 10% TRUE and 90% FALSE queried for TRUE values using an index takes 0.04 sec, while not using an index takes 0.24 sec. The index speeds up the query by about a factor of six.

I’ve repeated the test with worse case scenario of a binary column split 50/50. 50% true, 50% false. The numbers are a bit less consistent, but generally around 0.16 for the indexed version and 0.24 for the unindexed version. Go indexes!

People don’t pay any attention to URLs. Recently, a highly intelligent friend of mine inadverently posted the following link on Facebook: https://www.givingwhatwecan.org/get-involved/how-rich-am-i/?country=NLD&income=56000&adults=1&children=0

URLs are the building blocks of the world wide web. When a website changes its URLs, you get those nasty 404 errors. It would be good if website maintainers paid a little more attention to that.

As a website user, when you’re sharing a URL, look at it briefly. Does it actually contain the information you want it to contain? Can other people view this URL? Pro tip: use incognito mode to find out. Does it include superfluous information you were not intending to share? E.g. when searching google, you can easily end up with URLs such as https://www.google.com/search?q=previous+search#q=current+search

Keep URLs simple & working. Look at them briefly when sharing. It’s no rocket science.

Not knowing the following table can result in rather pointless arguments.

One day I’ll make a compendium of Polish-Czech insanities.
Nůžky ~ nożyczki, nožičky ~ nóżki.

RoboZZle is a robot programming game. You can play it in your web browser (even without Silverlight), on Android, or iPhone/Pad/Pod. Go and try it, otherwise the rest of this entry won’t make much sense.

After a while of playing the game myself, I started getting interested in creating a program to find puzzle solutions. Having completed Intro to AI by Sebastian Thrun and Peter Norvig in 2011 and never using the techniques since, this seemed like a good opportunity to explore them further. As things go, none of these techniques appear directly applicable and a simple evolutionary search seems best.

Approach

I started by writing a solution runner, the main output of which is determining whether a given RoboZZle program solves a puzzle. Next were a robozzle.com client, utilizing the awesome Python requests library, and finally solver randomly generating programs to be tested by the runner. You can browse the source code of Zlej Rob at github.

Some facts about the solver:

1. Start with an empty program.
2. In each generation, each of the surviving programs will have ~100 randomly generated offsprings.
3. A program is mutated by overriding (or removing) 1 to ~3 instructions of its parent.
4. Programs are scored: positive points for reaching a square, extra points for collecting stars, and a negative point for each program instruction.
5. If an offspring achieved higher score than its parent, it’s added to the current program population.
6. Program population is kept to ~100 programs with highest score.

I decided not to cross-breed programs at all, as it doesn’t feel like it would be helpful.

Surprisingly, keeping a set of all evaluated programs doesn’t eat all the memory (and prevents recalculating the same thing over and over, speeding up execution by an order of magnitude). Yay for sets of tuples!

Results

Zlej Rob has solved over a 1000 puzzles in a couple of weeks of running on my $5/mo DigitalOcean (that’s a referral link – sign up and I’ll get rich) droplet alongside this blog and a couple other things.

Zlej Rob’s good at solving one-way street puzzles with one function, bad at anything involving multiple functions, mediocre at multiple-possibilities puzzles, and passable at random walks.

Zlej Rob discovered the shortest solution for Twins 2 (which I improved by removing a redundant function call). I think that’s pretty impressive. The solution includes a lot of recursive calls, and the replay takes ages – the robot loops and loops, seemingly never getting very far.

Ideas for improvement

1. Visualisation. Would help identifying why certain things work and others don’t. I’ve almost started writing an Angular app to do that.
2. Smarter mutations. Mutations should include abstracting random instruction subchains into other functions (this could be very useful for multifunction problems where Zlej Rob usually fails terribly because it can’t connect the dots). It might also be better to insert instructions instead of replacing them.
3. Program diversity (“similarity penalty”). Surviving population for each generation is around 100 programs (toyed with higher values with no positive results – if anything, having 1000 programs is terribly slow), and they can all easily end up being very similar, getting stuck in local minima.
4. Higher score programs should have more offsprings. Could speed up certain puzzles but perhaps make less obvious solutions unatainable. Would need to ensure program diversity first.
5. Hints. For puzzles with an “obvious” order in which squares must be visited (such as binary counting or limit your stack), mandate this order. Currently, Zlej Rob finds random solutions that sweep vast majority of the board in a haphazard manner. Forcing the order in which squares must be visited would be extremely helpful, but likely requires human participation. Perhaps that’d be cheating?

Where do we go from here?

Zlej Rob’s results have surpassed my expectations, especially considering I haven’t spent that much time on it. Getting Zlej Rob into top 10 of RoboZZle players seems possible but would likely require an order of magnitude more effort than I spent so far. Not sure if worth it?