The stack powering tny.im: goodbye redundancy! The end of an era

I was casually going through my GitHub repos and came across PicoRed, a server redundancy manager I developed, with the immediate goal of managing the DNS records for the tny.im domain. The tny.im shortener used to be hosted by multiple servers in a Round-robin DNS configuration. The idea was that as servers went online and offline (or underwent maintenance, etc.), the DNS records would be automatically updated to reflect which servers are currently serving a service, in this case tny.im.

PicoRed is the successor to mersit, which served the same purpose but was written in very unidiomatic Python and was much clunkier than PicoRed (which is written in very unidiomatic Go, but used fewer resources and was somehow more stable). PicoRed and mersit were completely peer-to-peer, and this is because I couldn’t afford to have a “master” server that was stable enough and which I could be sure would have three nines of uptime.

The idea behind those tools is everything but novel; container orchestration, for example, requires similar tools to be deployed. For some reason, perhaps ignorance, back then I decided to write my own. (For an example of mersit/PicoRed done right, see Serf). I don’t regret it, of course: I learned a lot about distributed systems, and while my terrible consensus “algorithms” (a complete joke) worked, they taught me why things like Paxos and Raft had to be invented. The main takeaway was, “it’s complicated”. So for PicoRed I decided to use a library by Hashicorp that handled the hard parts for me (and that’s how my unidiomatic Go program was “somehow more stable”).

Three paragraphs into this post, and I’m still writing the introduction… these three paragraphs about distributed systems are just warming up for what’s coming, which is me saying that none of those homemade tools are in use anymore, and it’s not even because I switched to something better: tny.im, and some other services of the TNY network, are now served by a single server.

How did we get here? Back in 2014, I was a huge proponent of distributing every single service across many cheap servers, instead of buying a proper, rock-solid, big and expensive server from a reliable company. In theory, horizontally scaling would let one handle big amounts of traffic and improve availability at the same price, or even less – sounds great, right? These strong opinions were backed by the issues I was having with my BlueVM server. But now, we’re back to zero redundancy… what changed?

Well, my opinion is still the same: I’ll take horizontal scaling over vertical scaling any day, and the more redundancy that’s fit to pay, the better. The problem is when horizontally scaling begins to hurt performance and reliability instead of helping it, and that’s exactly what was happening in our case.

tny.im, dotAccount, PrizmID and my WordPress websites (this blog and the TNY network website) are powered by an extremely uninteresting LEMP stack. A LEMP stack is one composed by Linux, Nginx, MariaDB and PHP, or in other words, a LAMP stack but with Nginx instead of Apache. Until a few weeks ago, the “M” in this stack had the peculiarity of actually being MariaDB configured in master-master replication mode. What this means is that MariaDB was running on multiple servers, managing the same databases, and whenever a change was made, it was propagated to all of the other servers in the cluster (up to a few weeks ago, two servers; at some point in the distant past, up to five servers were used).

That’s how tny.im was served by multiple servers: simply by running the same PHP code in all servers, and having that code talk to the same database, replicated across all the servers. Of course, MariaDB master-master replication has its disadvantages. For one, performance is worse, because all database writes involve communication between the different MariaDB servers. This began to show on more database-intensive applications like dotAccount.

Perhaps more surprisingly, reliability is also worse. Perhaps I didn’t have MariaDB replication properly configured (after many attempts and hours spent, trust me), but it would sometimes break in wonderful states such as “WSREP has not yet prepared node for application use” whenever there was some network hiccup. This could happen as often as once a day, or once every two months (yes, networks are unpredictable like that). Whenever it broke, it would need to be manually restarted, and it would sometimes take multiple attempts until all the servers had their MariaDB running. In other words, exactly the opposite you want for a reliable system that requires minimum amounts of human supervision.

Perhaps PicoRed could have expanded into taking care of restarting the cluster, but since I couldn’t even get to a sequence of commands that, when executed on all servers at the right times, would reliably restart the MariaDB cluster, I kind of gave up. Lack of time and more interesting projects to develop, like Clouttery, meant that some stuff would inevitably get left behind, and my horrible mess of code called PicoRed ended up forgotten and eternally unfinished. Moving to proper solutions like Serf also required time that I didn’t have.

A few months ago I was notified that the provider of one of my VPS was closing, and all servers would be shut down by December 4th. I bought a new server, moved the stuff that wasn’t hosted anywhere else to it, but I really didn’t feel like reconfiguring the MariaDB cluster and PicoRed for the new server. PicoRed, in fact, stopped working in one of my servers (the one that wasn’t getting shut down) with some binary incompatibility error, a year or so ago. So I kind of gave up… reconfigured MariaDB so it stopped being a cluster, got rid of PicoRed, and said goodbye to one of the servers.

The new server is PHP and MariaDB/MySQL-free, and this probably won’t change. I would really like to move on from PHP and MariaDB to better languages and DBMSs. My main conclusion from the whole replication story is that MariaDB is not really prepared to scale horizontally, at least not without a lot of effort and “baby-sitting”.

I certainly have not given up on horizontal scaling, but I think that from on now, it’s best that I manage scaling at the application level instead of the database level, or alternatively, use a DBMS that was designed with horizontal scaling in mind, from the start. For the second option, it’s unfortunate that both CockroachDB and TiDB are still in a very premature state for production use.

I would rather not give up on relational databases; while it’s true that other types of database also cover some of the use cases of relational ones, I’m yet to know of any problem other than document storing that can’t be effectively solved with relational databases. (And for document storing, may I interest you in a relational database coupled with this strange thing called a filesystem?) Commercial solutions are obviously out of reach for me: it’s not like Segvault is a money-making machine; tny.im isn’t even profitable, despite all the ads!

I have grown to hate the mess of PHP and SQL that is tny.im so much, that shutting down the service (or at least getting it into a “read-only” mode) was once a topic for discussion at one of the TNY network meetings – all three of them. By the way, Segvault/TNY network is “hiring”, i.e. looking for new members with exciting project ideas, and if you had the patience to read this post this far, you may be a good candidate – contact me somehow.

Ads at tny.im earn me pocket change, that is used to offset the cost of the servers and domain names, and this shall be enough motivation to keep maintaining tny.im and supporting its users for some more years, updating MariaDB one version at a time.

My programming experience

Go to the bottom, “Summing it up”, for the TL;DR.

The day I turn this website into a portfolio/CV-like thing will come sooner or later, and arguably that’s a better use for the domain gbl08ma.com than this blog with posts nobody cares about – except when I rant about new operating systems from Microsoft. But if you really care about such posts, do not worry: the blog will still exist, it just won’t be as prominent.

Meanwhile, and off-topic intro aside, the content usually seen on such presentation websites everyone-and-their-cat seems to have these days, will have to wait. In anticipation for that kind of stuff, let’s go in a kind of depressing journey through my eight years programming experience.

The start

The beginning was what many people would consider a horror movie: programming in Visual Basic for Applications in Excel spreadsheets, or VBA for short. This is (or was, at the time; I have no idea how it is now) more or less a stripped down version of VB 6 that runs inside Microsoft Office and does not produce stand-alone executables. Everything lives inside Office documents.

Screenshot (30)

It still exists – just press Alt+F11 in any Office window. Also, the designer has Windows 7 Basic window styles… on Windows 10, which supposedly ditched all that?

I was introduced to it by my father, who knows his way around Excel pretty well (much better than I will probably ever will, especially as I have little interest). My temporal memory is quite fuzzy and I don’t have file timestamps with me for checking, so I was either 9, 10 or 11 years old at the time, but I’m more inclined to think 9-10. I actually went quite far with it, developing a Excel-backed POS system with support for costumer- and operator-facing character LCD screens and, if I remember correctly, support for discounts and loyalty cards (or at least the beginnings of it).

Some of my favorite things I did with VBA, consisted in making it do things it was not really designed for, such as messing with random ActiveX controls and making it draw strange-looking windows (forms) and controls through convoluted Win32 API calls I’d have copied from some website. I did not have administrator rights to my computer at the time, so I couldn’t just install something better. And I doubt my Pentium III-powered computer, already ancient at the time (but which still works today), would keep up with a better IDE.

I shall try to read these backup CDs and DVDs one day, for a big trip down the memory lane.

Programming newb v2

When I was 11 or 12 I was given a new computer. Dual core Intel woo! This and 2GB RAM meant I could finally run virtual machines and so I was put on probation: I administered the virtual computers, and soon the real hardware followed (the fact that people were tired of answering Vista’s UAC prompts also helped, I think). My first encounter with Linux (and a bunch other more obscure OS I tried for fun) was around this time. (But it would take some years for me to stop using Windows primarily.)

Around this time, Microsoft released the Express (free) editions of VS 2008. I finally “upgraded” to VB.NET, woo! So many new things to learn! Much of my VBA code needed changes. VB.Net really is a better VB, and thank Microsoft for that, otherwise the VB trauma would be much worse and I would not be the programmer I am today. I learned much about the .NET framework and Visual Studio with VB.NET, knowledge that would be useful years later, as my more skilled self did more serious stuff in C#.

In VB.NET, I wrote many lines of mostly shoddy code. Much of that never saw the light of day, but there are some exceptions: multiple versions of Goona Browser made their way to the public. This was a dual-engine web browser with advanced UI, and futuristic concepts some major players copied, years later.

How things looked like, in good days. Note the giant walls of broken English. I felt like "explain ALL the things"!

How things looked like, in good days (i.e. when it didn’t crash). Note the giant walls of broken English. I felt like “explain ALL the things”! And in case you noticed the watermark: yes, it was actually published to Softpedia.

If you search for it now, you can still find it, along with its website which I made mostly from scratch. All of this accompanied by my hilariously broken English, making the trip to the past worth its weight in laughs. Obviously I do not recommend installing the extremely buggy software, which, I found out recently, crashes on every launch but the first one.

Towards the later part of my VB.NET era, I also played a bit with C#. I had convinced myself I wanted to write an operating system, and at the time there was a project called COSMOS that allowed for writing (pretty limited) OS with C#… of course my “operating” systems were not much beyond a fancy command line prompt and help command. All of that is, too, stored in optical media, somewhere… and perhaps in the disk of said dual-core computer. I also studied and modified open source programs made in C# (such as the file downloader described in the Goona Browser screenshot) for my own amusement.

All this happened while I developed some static websites using Visual Web Developer Express as editor. You definitely don’t want to see those (mostly never published) websites, but they were detrimental to learning a fair bit of HTML and CSS. Before Web Developer I had also experimented with Dreamweaver 8 (yes, it was already old back then) and tried my hand at animation with Flash 8 (actually I had much more fun using it to disassemble existing SWFs).

Penguin programmer

At this point I was 13 or so, had my first contact with Linux more than done, through VMs and Live CDs, aaand it happened: Ubuntu became my main OS. Microsoft “jail” no more (if only I knew what a real jailed platform was at the time…). No more clunky .NET! I was fed up with the high RAM usage of Goona Browser, and bugs I was having a hard time debugging, due to the general code clumsiness.

How Ubuntu looked like when I first tried it. Good times. Canonical, what did you do?

How Ubuntu looked like when I first tried it. Good times. Canonical, what did you do?

For a couple of years, in terms of desktop development, I only made some Python scripts for my own amusement and played a very small bit with MonoDevelop every time I missed .NET. I also made a couple Lua scripts for Rockbox. I learned much about Linux usage and system maintenance as I used it more and more on my own computers and on my first Virtual Private Servers, which I got after much drama in the free web hosting communities. Ugh, how I hate CPanel.

It was around this time that g.ro.lt and n.irc.su appeared. g.ro.lt was a URL shortener that would later evolve into 4.l.to and later tny.im. n.irc.su was a social network built on Elgg, which obviously failed. I also made some smaller websites, like one that would take you to random image hosting websites, URL shorteners and pastebins, so you would not use the same service every time you urgently needed one. These represented my first experiences with PHP programming.

I have no pictures to show. The websites are long gone, not on the Internet Archive, and if I took screenshots, I have no idea where I put them. Ditto for the logos. I believe I still have the source code for the random-web-service website somewhere, at least the front page layout.

All this working on top of free stuff: free (and crappy) subdomains, free (and crappy) web hosting, free (and less crappy) virtual servers. It would take me some time until I finally convinced myself I needed to spend some money for better reliability, a gist of support and less community drama. And even then I would spend Bitcoin, which I earned back when it was really cheap, making the rounds of silly faucets and pulling money out of CPAlead-like offers through the use of multiple proxies (oh, the joy of having multiple VPS…). To this day I still don’t have a PayPal account.

This time, and when I actively developed tny.im (as opposed to just helping maintain it), was the peak of my gbl08ma-as-web-developer phase. As I entered and went through high school, I would get more and more away from HTML and friends (but not server maintenance), to embrace something completely different…

Low level, little resources: embedded systems

For high school math everyone had to use a graphing calculator. My math teacher recommended (out of any interest) Casio calculators because of their ease of use (and even excitedly mentioned, Casio leaflet in hand, the existence of a new and awesome color screen model that “did everything and some more”). And some days later I had said model in my hands, a Casio fx-CG 20, or Prizm, which had been released about a year before. The price difference from the earlier dot-matrix screen Casio calcs was too small to let the color screen go.

I was turning 15, or had just turned 15. I remember setting up the calculator and thinking, not much after, “I want to code for this thing”. Casio’s built-in Basic dialect is way too limited (and after having coded in “real” languages, Basic was silly). This was in September 2011; in March next year I would be releasing my first Prizm add-in, CGlock, a calculator PIN-locking software.

Minimalist look, yay! So much you don’t even notice it’s a color screen.

This was my first experience with C; I remember struggling with pointers, and getting lots of compilation warnings and errors, and run-time errors. Then at some point everything just “clicked in” and C soon became my main language. Alas, for developing native software for the Prizm, this is the only option (besides using C++ without most of its features, not even the “new” keyword).

The Prizm is a horrible platform, especially for newbie C programmers. You can’t use a debugger, nor look at memory contents, the OS malloc/free implementation has bugs (and the heap is incredibly small, compared to the stack) and there’s always that small chance some program damages your calculator, or at least corrupts your estimated files and notes. To this day, using valgrind and gdb on the desktop feels to me as science fiction made true. The use of alloca (stack allocation) ends up being preferred in relation to dynamic allocation, leading to awkward design decisions.

Example of all the information you can get about an error in a Prizm add-in. It’s up to you to go through your binary (and in some cases, disassemble the OS) to find out what these mean. Oh, the bug only manifests itself when compiling with optimizations and without symbols? Good luck…

There is a proprietary emulator, but it wasn’t designed for software development and can’t emulate certain things. At least it’s better than risking damage to expensive hardware. The SuperH-4 CPU runs at 58 MHz and add-ins have access to about 600 KiB of memory, which is definitely better than with classic z80-powered Texas Instruments calculators, but one still can’t afford memory- or CPU-intensive stuff. But what you gain in performance and screen resolution, you lose in control over the hardware and the OS, which still have lots of unknowns.

Programming for the Prizm taught me how it’s like to work without the help of the C standard libraries (or better, with the help of incomplete and buggy standard libraries), what a stack overflow looks like (when there’s no stack protection), how flash memories work, what DMA is, what MMUs do and how systems can be bricked when their only bootloader is not read-only. It taught me how compilers work from an end-user perspective, what kind of problems and advantages optimizations introduce, and what it’s like to develop parts of the C standard library.

It also taught me Casio support in Portugal (Ename) is pretty incompetent at fixing calculators, turning my CG 20 into a CG 10 and leaving two big capacitors out of a replacement main board. In this hardware topic, I learned quite a bit about digital logic from Prizm hardware discussions at Cemetech. And I had some contact with SH4 assembly and a glimpse into how to use IDA Pro. Thank you Casio for developing a system that works so well and yet is so broken in so many under-the-hood ways, and thank you Cemetech for briefly holding the Prizm higher than TI calcs.

I developed other add-ins, some from scratch and others as ports of existing PC software (such as Eigenmath). I still develop for the Prizm from time to time, but I have less and less motivation as the homebrew community has stagnated and I use my Prizm much less, as I went to university. Experience in obscure calculator platforms does not make for a nice CV.

Yes, in three years or so I went from the likes of Visual Studio to a platform where the only way to debug is to write text to the screen. I still like embedded and real-time programming a lot and have moved to programming more generic and well-known things such as the ESP8266.

Getting in the elevator

During the later part of high school (which I started in the fall of 2011 and ended in the summer of 2014), I did more serious Python stuff, namely Mersit, later deprecated in favor of Picored, which is not written in Python but in Go. Yes, I began trying higher-level stuff again (higher level, getting in the elevator… sorry, I’m bad at jokes).

My first contact with Go was when I was 17, because I wanted to develop something that ran without external dependencies (i.e., unlike Java or .NET) and compiled to native code. I wanted to avoid C/C++, but I wasn’t looking for “a better C” either, so Rust was not it. Seeing so much stuff about Go at Hacker News, one day I decided to try my hand at it and I like it quite a lot – I’m still unsure if I like it because of the language itself or because of the great libraries one can use with it, but I think both play an important role.

GolangPicoRed

This summer I decided to give C# another chance and I’m quite impressed – turns out I like it much more than I thought. It may have something to do with trying it after learning proper languages vs. trying it when one only knows VB. I guess my VB.NET scars are healed. I also tried a bit of Java, in my first contact with it ever, and it seems my .NET hate converted into Android API hate.

Programming with grades

University gave the opportunity (or better, the obligation) of having other people criticize my code. The general public could already see the open-source C code of my Casio Prizm add-ins, and even the ugly code of Goona Browser, but this time my code was getting graded. It went better than I initially thought – I guess the years of experience programming in different languages helped, especially as many of the people I’m being compared with have only started programming this year.

In the first semester we took an introductory programming course, which used Python, and while it was quite easy for me, I took the opportunity to learn Python to a greater depth than “language in which to write quick and dirty glue code”. You see, until then I had not used classes in my Python code, for example. (This only goes to show Python is a versatile language, even if slow.)

We also took an introductory computer architecture course where we learned how basic CPUs work (it was good for gluing all the separate knowledge I already had about it) and programmed in assembly for a course-specifc CISC-like architecture. My previous experience with reading SH4 assembly proved quite useful (and it seems that nowadays the line between RISC and CISC is more blurred than ever).

In the second semester, I had the opportunity to exercise my C knowledge, this time not limited to the Prizm platform. More interestingly, logic programming, a paradigm I had no intention of ever programming in, was presented to us. So Prolog it was. It went much better than I anticipated, but as most other people who (are forced to) learn it, I have no real use for it. So the knowledge is there, waiting for The Right Problems(tm). I am afraid I’ll forget much of it before it becomes useful, but if there’s something picking C# up again taught me, is that I can pick up pretty fast skills learned and abandoned long ago.

The second year is about to begin and there’s some object-oriented programming coming, I hope I do well.

Summing it up

I have written non-trivial amounts of code in at least 8 languages: Visual Basic, PHP, C#, Python, Lua, C, Go, Java and Prolog. I have contacted with two assembly dialects and designed web pages with HTML, CSS and Javascript, and of course automated some tasks with bash or plain shell scripting. As can be seen, I’m yet to do any kind of functional programming.

I do not like “years of experience” as a way to measure language proficiency, especially when such languages are learned for use in short-lived side projects, so here’s a list with an approximate number of lines of code I have written in each language.

  • C: anywhere between 40K lines and 50K lines. Call it three years experience if you will. Most of these were for Prizm add-ins, and have since been rewritten or heavily optimized. This is changing as I develop less and less for the Prizm.
  • PHP: over 15K lines, two years if you want to think that way. The biggest chunk of these were for developing the additions to YOURLS used in tny.im, but every other small project takes its own 200-500 lines of code. Unfortunately, most of this is “bad” code, far from idiomatic. The usual PHP mess, you know.
  • Python: at least 5K lines over what amounts to about six months. Of these, most of the “clean” lines (25-35%) were for university projects.
  • Go: around 7K lines, six months. Not exactly idiomatic code, but it’s clean and works well.
  • VBA: uh, perhaps 3 or 4K lines, all bad code 🙂
  • VB.NET: 10K lines or so, most of it shoddy code with lots of Try…Catch to “fix” the problems. Call it two years experience.
  • C#: 10K lines of mostly clean and documented code. One month or so 🙂
  • Lua: mostly small glue scripts for my own amusement, plus some more lines for use in games such as Minetest, I estimate 3-4 K lines of varying quality.
  • Java: I just started, and mostly ported C# code… uh, one week and 1.5K lines?
  • HTML, CSS and JS: my experience with JS doesn’t go much beyond what’s needed to modify DOM elements and make simple AJAX requests. I’ve made the frontend for over 5 websites, using the Bootstrap and INK frameworks.
  • Prolog: a single university assignment, ~250 lines or one month. A++ impression, would repeat – I just don’t see what for.

In addition to all this, I have some experience launching the programs and services I make – designing logos/branding, versioning, keeping changelogs, update instructions, publishing, advertising, user support. Note that I didn’t say I’m good at any of these things, only that I have experience doing them, for better or worse…

Things I’d like to have more experience with:

  • Continuous integration / testing in general;
  • Debugging code outside of .NET/Visual Studio and printing debug lines in C;
  • Using Git and other VCS in big repos/repos with more people (I want to see those merge conflicts and commits to the wrong branch coming);
  • Server-side web development on something other than PHP and Go. And learning to use MVC frameworks, independently of the language;
  • C++ (and Java, out of necessity. Damned Android);
  • Game development. Actually, this is how many people start, but I’m so cool that I started by developing POS software 🙂

Distributed systems and mersit, a Tiny Server Redundancy Manager

My previous post on this blog was published by the end of the long-gone month of June. Many things have changed since then, for example, I entered university and was pressed into creating a Facebook account (more or less separate from the rest of my online presence, so don’t look for me, I won’t add you). On that post, I rambled about the recovery from a big server outage that costed 42 hours of tny.im downtime, and over one week of server downtime. I learned my lessons (I doubt BlueVM learned theirs, but that’s a whole other story), and I went forward with what I said I would do: “setting up a new advanced and redundant system” for ensuring tny.im is always up.

That system has been up and running for over two months now, with varying amounts of servers making the redundancy and load balancing, and a plethora of occasional hiccups. Right now it’s composed of three virtual servers (all from different providers…), but there were times when it was composed of five servers. These three servers are paid, and while they aren’t exactly expensive (but not the cheapest, either), you can imagine the bill, so let’s not talk about tny.im profitability now, OK? (I have kind of given up).

In the spirit of the great statisticians of our time, here's a graph without title, labels or axis.

In the spirit of the great statisticians of our time, here’s a graph without title, labels or axis.

However, having three servers serving the same website, with all three of them being almost a clone of each other (which means, all have the same files and database contents, synced), in a DNS round-robin setup doesn’t directly lead to greater uptime. In fact, I have found out it can lead to more outages, since now the total downtime is approximately the sum of the downtime of each server. Of course, most of these outages are partial (as in, only users unlucky enough to have their DNS request resolved to the IP of a server that is down, will actually perceive the site as down), except for when the MariaDB replication freaks out and basically grinds all database operations, on all servers, to a halt, requiring a complicated manual restart of all MariaDB instances, in a specific order (yes, I have spent many hours searching for an alternative database system, and couldn’t find any that met my requirements).

In order to actually achieve greater uptime, one must have a system that automatically manages the DNS records so that the domain(s) of the website in question never have any records pointing to servers that are down. In other words, the “sheep” must be “hidden from sight” as soon as they go “bad”, and should be put back “in stage” once they become “good”. Being DNS something that was definitely not made for real-time record edits, with many systems caching DNS request results well beyond the specified TTL, this system obviously doesn’t ensure that the “bad sheep” are not invisible to everyone watching the show. But if it manages to do it for even a small percentage of the public, it’s already better than not hiding from anyone (and especially, if it successfully hides the problem from the uptime monitor, that’s even better 🙂 ). This explains why the DNS records for tny.im are set with TTLs of five minutes.

The development of such a DNS record management system was also more or less contemplated in my previous post, when I say:

I’ll take this downtime and new server acquisition as the motivation for setting up a new advanced and redundant system, so that if one server goes down, tny.im (and possibly this blog too) will continue to operate as normal.

And in the end, in a later edit:

On related news, Mirasm – the Tiny Server Redundancy Manager – is mostly finished, only needs some more testing to be put on production servers, managing the new tny.im redundancy system.

“Mostly finished”, as we all know, really means “It’s 99% ready, I only need to figure out the remaining 1% that consists on… everything that is tricky and I’m not sure how it’s done”. This is specially true in this case, as I had high requirements for my manager: it couldn’t use any resources other than the servers I had already (it would’ve been easy to have a separate server just for monitoring and editing DNS as needed, but I didn’t want to pay for yet another server on yet another provider), and it couldn’t fail more than tny.im itself. In fact, the time when the manager has to do more important work, is when it is not working, i.e. when a server goes down and so goes the manager. I finally finished the project, and it works as planned. I only got the name wrong…

Introducing mersit, a Tiny Server Redundancy Manager

Pronounced “m-eh-rs-ee-t”, with the first “e” being like the one in “explain”, mersit is a simple Python script (Python 2.7, because I wasn’t sure what libraries were available for 3.x nor if my servers would run it well) hacked together with some sections that definitely look like spaghetti code. The good news is, it works fine, and has been well tested, so if you study it in the “black box” way, there are no big problems with it.

The purpose of the script is to manage the DNS records of the website served by the group of synced servers, in this case, tny.im. It runs on each server, in a peer-to-peer fashion. The peers select a single master, that will monitor all the peers and manage the DNS as they go up and down, “deciding who’s on stage”, and all peers will check whether the master is up, and select a new one that will edit the DNS to “hide the master from the public” when it goes down.

I definitely want to open-source mersit at some point, but not now because it’s not ready for prime-time (see “spaghetti code”, above), and I want to change some things that will make it more general-purpose. mersit has been managing the live records for tny.im for the past week (it’s been peaceful).

Continuing our journey through the world of meaningful graphs, here's another.

Continuing our journey through the world of meaningful graphs, here’s another.

I have gone so far as to write a read-me for mersit (mainly for me to read, as I know I’ll forget how it works within six months). I think it’s best if I put the start of the read-me here, instead of trying to explain it all, once again:

mersit - Tiny Redundant Server Manager
Copyright 2014 tny. internet media
This version is customized for tny.im

This is a Python 2 script that manages a group of computers/servers/thin clients/machines in a network (local- or wide-area), by automatically executing actions when something relevant happens to one of the machines.

We'll call the "machines" "peers". mersit assumes all peers and the network are trusted.

The script is meant to be run directly on the peers that are to be controlled, in a setup where there is not a single point of failure. It is not of much use when run in a single peer; in the context of this script, a "group" only starts to make sense when it has over one element.

We'll refer to this script as "controller software" or simply "controller", and to the other software that runs on a peer and which is to be monitored as "application". The controller is made to run unattended, even though it accepts commands (issued by an "operator") to trigger certain behavior manually.

The "something relevant" mentioned in the first paragraph consists on one of these "events of interest":

- A peer goes "online", that is, it is reachable by other peers and reports the status of its controller software as "OK" or "ready";

- A peer goes "offline", that is, it is either not reachable by at least some peers, or the controller is reporting its state as "not good" or "not ready";

- A peer becomes good-for-work (GFW), which means, that the application is functioning properly and performing its function (such as listening for incoming connections, data to process, etc.);

- A peer becomes not-good-for-work (NFW), in which case the application is not functioning properly, is too busy to perform its function (over capacity), or is otherwise unavailable.

Each peer works in a given "domain", which is the group the peer belongs to. The domain is specified by a name and secret which act basically like a username and password pair. Peers will only communicate with other peers of the same domain, that is, peers where the domain name and secret are the ones the controller is configured to use. The domain acts as the authentication element; an external party can not join, communicate or perform actions in a domain unless it knows the name and password used by the peers of the domain.

(Please note, that communication between peers is not encrypted by the controller - it goes completely plain-text over the network. It is possible to secure the communication between peers using external tools; such secure functionality goes beyond the scope of this software. The "domain" is simply a basic authentication system, implemented using HTTP authentication, to ensure that peers of a certain group don't start talking with peers from other groups. The basic authentication system is enough to protect against the casual script-kiddie, but by no means adequate for protection from a malicious party in an untrusted/open network)

The controller on each peer must know _a priori_ (i.e. before it starts) about where to find at least some of the controllers on other peers. Peer discovery doesn't happen automatically, however, once a peer's controller can communicate with another controller, it will add every controller in the "contact list" of the latter to its "contact list".

Imagine the following situation: you have peers A, B and C (and their controller software). The controller in A only knows about peer B. The controller in B only knows about peer C. If you start the controller on peer A, then start the controller on peer B, peer A will tell peer B about its existence, and peer B will tell peer A about the existence of a peer C (independently of peer C being running/reachable). However, if the controller in A knew about no peer (other than itself), it would never find peer B or C even if their domain settings all matched. Even though a big domain can be bootstrapped from just two peers, to ensure good operation, all controllers should know about all peers. This way, if the controller on a peer resets for some reason, it will have a greater chance of reaching another peer.

The "contact list" is the list of "known" peers. The controller keeps three lists of peers in memory: the "known" peers, the "reachable" peers, and the "GFW" (good-for-work) peers. The list of known peers is initialized from the source code's configuration section when the controller starts. It then proceeds to see which peers are "reachable", that is, can be reached through the network, are in the same domain (not being in the same domain gives the same effect as not being reachable over the network) and have their controller software report its state as "OK".
This initial status checking includes the exchange of some other information about the controller. Once this initial peer identification is done, the controller enters a monitoring loop where it will keep the contents of the three lists up-to-date. The controller keeps running this infinite loop throughout most of its lifetime. How the lists are kept up-to-date and what happens when their contents change is something that depends on the current controller mode.

There are two possible modes for controller operation: master and non-master. There is exactly one controller in master mode per domain, and this controller is usually called "the master" (the master peer has the controller in master mode). The differences between the modes are mostly related to what happens in the monitoring loop, but before going into those differences, it is important to understand how the controllers decide which peer is the master peer.

When a controller starts and there are no reachable peers, it promotes itself to master, since there must be exactly one master per domain. Later, when another controller joins the domain (either because it started or because it went online after e.g. a period without connections or power), it checks which peers are reachable from its "known" list and "asks" them which is the master peer. Every peer should reply with the same peer, in which case the new controller assumes that peer is the master, and informs the master about its existence, to account for the fact that the new peer may not be in the master's "known" list.

However, and especially on domains where not all peers initially know about every other peer, it's possible that a "head split" occurs and there are two masters in the same domain. Imagine a domain where there are four peers D, E, F and G. D only knows about E, which in turn only knows about D. F doesn't know about any peer, and we'll leave G aside for now. All peers are offline.
The D controller starts up, sees it can't reach the only peer it knows (E), so calls itself master. The E controller starts up and reaches D, D says it is the master, E assumes D is master, all is fine.
The F controller starts up, sees it can't reach any peer because its "known" list is empty, so calls itself master and sits quietly waiting for someone to contact it, which in turn would let it know about more peers.
We now have the following situation ([M] represents a controller in master mode, --- represents the knowledge peers have of each other):

 -DOMAIN------------------
|                         |
|   D[M]-----E     F[M]   |
|                         |
 -------------------------

Things could be like this forever, and no conflicts would occur - however, this is probably not a domain you want to have, since F doesn't know about any "event of interest" related to D or E, and these two don't know about any events related to F. In this situation, D--E and F act like separate domains.
Assume that G is a peer which knows about D, E and F, and that its controller starts up, contacting D, E and F. The first two will agree that D is the current master, but F will disagree and say it is the master. At this point we have a conflict. There are many ways to solve this, including some form of "voting" (e.g. the peer the largest amount of the peers say is the master effectively becomes it), but mersit solves this in a simpler way.

The controller checks that everyone in the domain agrees on what peer is the master on every iteration of the monitoring loop. It does this by "asking" each peer in the list of known peers who is the master. The first peer asked is free to reply with any peer. The ones that are asked next must agree with the first one. If not, the controller that was doing the loop tells each disagreeing peer that the actual master, is the one from the first peer's reply. It is possible that a minority is asked first, and thus everyone is forced to "change its opinion" to that of the minority. This is not a problem - mersit assumes all peers are trusted. Note that it can sometimes take some iterations of the monitoring loop for all peers to settle on a single master, because two (or more) peers may be trying to "change the opinion" of the other peers to different masters. This is not a problem either, because even if this kind of concurrency conflict happens once or twice in a row, it will stop happening as soon as one peer is faster than the other to tell everyone (including the other peer(s) that are trying to "change opinions"). What matters is that in the end, every peer knows about all others, and there is a single master. In this case, it could be D:

 -DOMAIN--------------------
|                           |
|  D[M]-----E-----F-----G   |
|                           |
 ---------------------------

If the master becomes unreachable, or its controller stops working, the other peers will also find themselves a new master, by sorting the list of reachable peers alphabetically and choosing the first peer in the sorted list. Of course, if for some reason the list is not consistent across peers, the peers will try to "convince" others to settle on who they "think" is the master as previously explained, until everyone is set to the same master.

Being the master essentially changes what happens in the monitoring loop. When a controller is in master mode, it is responsible for updating the list of "reachable" and "GFW" peers, by checking which peers are reachable (both in terms of network and in terms of functioning controller) and which have the application in a working condition. If there are changes in the lists that indicate an event of interest, it runs the appropriate handler. If, for example, a peer becomes NFW due to a problem in the application, it will stop being in the GFW list, and the handler function for when a peer leaves that list will be run with the peer in question as the argument. If the master becomes unreachable (network error, controller error, etc.), a new master will be found, as explained in the previous paragraph, and the new master is responsible for running the handler with the previous master as argument.

When a peer is not master, it won't run any handlers for events of interest, and it is not responsible for updating the "reachable" and "GFW" lists - it will retrieve these from the master. The controllers on all peers need to keep their lists up-to-date, sharing a "vision of the domain" similar to that of the master, so that any peer can become a master instantly in case of necessity, without having to spend time performing checks on all peers and ensuring it has the best-and-latest list of "known" peers.

The operator can manually tell a controller to become the domain's master. When the appropriate command is issued, the controller will send a command to every other controller instructing them to switch to the new master. This command may not always have an effect in some controllers, because while the first controller is sending the commands, other controllers are seeing if everyone agrees on who's the master, and issuing the same commands with another master in mind. This is a sequence of events picturing the situation, in a domain where there are three peers H, I and J, and H is the initial master:

 0. ...
 1. Peer H checks that every controller agrees it is the master (all agree);
 2. Peer I checks that every controller agrees H is the master (all agree);
 3. Peer J checks that every controller agrees H is the master (all agree);
 4. Operator issues command for peer I to become master;
 5. Controller on I assumes it is master, starts sending commands to other peers;
 6. Peer H checks that every controller agrees it is the master, before the message from I that I is the new master can get to H;
 7. Peer H finds out I (and possibly others) don't agree, sends them commands to change the master back to H;
 8. Peer I changes master back to H;
 9. Peer I checks that every controller agrees H is the master (all agree);
 10. Peer J checks that every controller agrees H is the master (all agree);
 11. ...
 
If the master doesn't change when the manual command is issued, it's a matter of trying again. Most often, this kind of concurrency problem does not occur, and even when it does, it does no damage. While it is true that mersit could detect this situation and keep issuing commands automatically until the decision takes effect, we chose to not make it this way to allow the human operator finer control.

The primary focus of mersit is to monitor a distributed application. The master checks if the application, or part of the application, running on a certain peer is in working condition by asking that peer's controller about the state of the application it is monitoring. In turn, this controller runs a function, defined by the mersit user in the mersit source code, that should check the application and return True (if application OK) or False (if not). This can involve, for example, making a HTTP request to a HTTP server in that peer to verify it is working. The controller then communicates the status of the application to the master (which may be itself). All this shouldn't take too long, especially when the domain has many servers, as only one peer is asked at a time. If checking the status of the application typically takes over one second, it is best to store the last known status in a variable, and update that state periodically in an asynchronous manner that may be external to the mersit script.

The part related to DNS records is not explained on the read-me, because it is related to the handlers (which each mersit user would customize to the specific needs of the system – as I said, I tried to make it a general-purpose script). Sounds interesting? Feel free to ask questions, or point out problems, in the comments.

Recovering from prolonged outage

tny.im and this website were down for about 42 hours, starting on June 29 at 03:16 UTC.

The problem? BlueVM’s S19-NY server went down, taking with it the server I have/had there (and which I paid for a full year!). Other than this outage, the service had worked fine for three months, – fast network, full resources availability – since I bought it.

S19-NY is still down, without any ETA for it to come back. There’s no information in what conditions it will come back, or *if* it will come back (with the previous contents, at least). BlueVM staff is pretty much unresponsive, other than a guy who sometimes hangs on IRC and says he can’t fix the KVM instance because he doesn’t have access to it.

Of course, I no longer recommend BlueVM and I don’t plan on renewing the server I have with them.

The “solution” to put an end to two days of downtime, was to buy the cheapest SecureDragon OpenVZ server, (OpenVZ! so hard to live without my beloved KVM, I can’t even use davfs2 because there’s no fuse module!) and restore the backups I had (from four hours before the BlueVM instance went down). This has been done, except for the HTTPS certificate of tny.im… that alone is another story:

As I tried to retrieve the existing cert from StartSSL (because, stupid me, automated backups were not copying everything SSL-related, and I didn’t save it locally), I found my authentication certificate had expired. This basically means my StartSSL account is lost, unless I create a new account and ask their staff to link it to the old one. They probably won’t do that without a payment and some ID checks so… out of question. I guess, if the BlueVM server doesn’t come back, that I’ll just create a new StartSSL account and generate a new cert for tny.im. There’s no security issue with this, as the previous certificate has not been compromised (unless BlueVM is collecting certificate private keys from inside their clients’ machines…) and so it doesn’t need to be revoked.

To conclude the HTTPS point: tny.im has the HTTPS service unavailable for now, until I can retrieve the existing certificate from the previous server, or until I get a new one.

Is all the fault in BlueVM’s side? Of course not… I could have lost my love for the money earlier, bought the SecureDragon VPS yesterday already and reduce the downtime by 24 hours. But since I had hope the problems on BlueVM support were just Sunday-related, I thought that by Monday they would have it fixed. They didn’t.

On related news, I’ll take this downtime and new server acquisition as the motivation for setting up a new advanced and redundant system, so that if one server goes down, tny.im (and possibly this blog too) will continue to operate as normal. I have two servers already (assuming the BlueVM one comes up), and I plan on developing a system where firing up new instances of tny.im on any empty server will be really easy. The system will be always prepared to lose any server at any point, without data loss, and restore full service within five minutes. That way, I can add less reliable hosts, perhaps even VPS trials, to the redundant system. This also allows for scaling the service as needed. Sounds ambitious? Of course it won’t happen in a week, but I have the full summer to develop and test the system…

Why don’t I just go with some SaaS that supports scaling? Two reasons: the price is too high, and the tny.im software is not coded in a way that’s compatible with these services. Let me remember you, that while not exactly being a CGI script, tny.im is not coded in one of those fancy modern languages, and even though PHP is not exactly outdated or unmaintained, the quality of the code can make it pretty horrible or pretty good. And the code is… not perfect – it doesn’t use any popular framework, it is based on YOURLS and has many, many hacks feature additions, plus a… very close relationship with the database.

Let me finish by saying that downtime of this kind is something to be expected if I were still hosted by FreeVPS and the like. But believe it or not, on FreeVPS and other sponsorships I’ve never seen a customer service as bad as the one of BlueVM (and it’s hard to remember an outage as big as this one). It is definitely not adequate for a paid service. In addition to S19-NY, they have many other instances down, with similar or worse downtime. The admins don’t appear to be online or reachable in any way, even by other staff members. The latest news/excuse on the S19-NY situation is that IPMI is broken and they are waiting for the provider to fix it… now tell me, does this look like a serious company, or some poor-man’s sponsorship?

EDIT: The BlueVM server is still down. tny.im is now hosted by three servers with round-robin balancing. HTTPS service was restored with a new certificate.

EDIT 7/7/2014: I forgot to update this post in time, but the BlueVM server has been up since three days ago. But I only got to know that the service was restored thanks to a friend of mine, because they didn’t reply to my ticket to inform me about it. Anyway, I don’t plan to renew this server, and BlueVM lost me as a customer (except for some really cheap deal which I’ll use as a personal/development box, and never in production).

On related news, Mirasm – the Tiny Server Redundancy Manager – is mostly finished, only needs some more testing to be put on production servers, managing the new tny.im redundancy system.

Server status updates

Just in case you didn’t know/notice, this website and tny.im are no longer hosted on a VPS provided by FreeVPS. Due to issues on the node where that VPS was hosted, that made it become network-unreachable, I needed to buy a new server to avoid a longer downtime.

I ended up buying a KVM-powered VPS from BlueVM because they accept Bitcoin and had a good special deal at the time. It’s my third day with the server and so far, so good. It has better specs than the server I had, and after config optimization tny.im is now able to handle more concurrent requests than it did before. I will keep readers updated and post a detailed review of the server, later.

Moving away from sponsorships is be better in the long term (less worries, proper support…), even if it means that tny.im is now unprofitable (temporarily, hopefully).

Crowdfunding

If I tried to crowdfund the renewal of my domain names or server hosting, I wonder how well that would work…

Server changes

I moved this blog to a different server (no downtime because the transfer was planned). Please report if something is broken.

On other topic, I have been very busy with school and real life in general, even on weekends; that sums up the reasons for my recent online absence.

Updates on lots of things

I figured out I didn’t post anything here for quite a while – actually more than two months! There have been a series of events that have stopped me from updating this blog and, for a good part of these two months, have stopped me from doing anything besides checking my email and doing some school work.

The first cause for my online absence was that, a few days after I posted about this server having been reloaded, school classes have started. I happen to be taking probably the hardest school year I ever took – at least, that’s what older people said before it started, and which proved to be true now that two months of classes have gone by. Free time isn’t that abundant and I prefer to spend it doing actually something more useful than writing about random things on this blog. That almost explains why I haven’t been active on communities like FreeVPS. But there’s more…

Some time in mid-October I was left without a reliable, 24/7 Internet connection. This, after two or three months without an unlimited-traffic Internet connection, during which I had to hold the downloading of system updates for both Linux and Windows on all of my machines, plus the Android tablet; to keep to a minimum the access to online streaming media; and of course to hold any other big download/upload tasks. I had no Internet connection at home, which means checking the emails at school when possible. Checking anything other than email is horribly slow and unpractical on my old phone, and that’s why I repeat email is the fastest and most reliable way of communicating with me.

At the same time, I was being kept busy with school work. Free time was spent developing an Utilities software for my graphic calculator, a Casio fx-CG 20; this was probably the most useful thing I could do without an Internet connection, because at least gcc still hasn’t moved to the cloud.

Around 1st November, the technical problems that were stopping me from having Internet at home were solved (read: free wifi hotspot magically came back online!), and so the second cause for my lack of participation in the online life was solved. But there was still the first cause: school. So slowly, as I was having free time, I began to put me up to date on what happened while I was offline, and to make slow progress on my projects.

Apart from all this, I also have friends and family, and my real-life social life is way more active than it used to be one or two years ago (my brain must have installed some system updates, finally).

To conclude, on the IT side of things I’m now developing a JPEG image viewer for the Casio Prizm, using the picojpeg library (incredibly useful). The thing already works but is damn slow, plus selecting files is still a pain; this Prizm add-in isn’t yet published anywhere, or even announced on any place except this blog post (and don’t consider this an official announcement).

I finally had time to fix some bugs at tny.im and add some features: the optional short URL toolbar is now powered by Meny, which was developed by a guy which also has lots of other awesome work, namely a promising online presentations service called rvl.io.

I’m also trying to keep participating on online communities such as FreeVPS and Cemetech, and to keep my Twitter feed rolling, but the first cause for my online absence is still valid, and will only stop being on the beginning of July next year. This doesn’t exactly mean I won’t be online, but that you shouldn’t expect me to be as active as I once used to be. Again, if you need something, email me (admin at tny dot im, or my username on Gmail).

Until next post!

Server reloaded

The dedicated server where the VPS that hosts this blog was installed had a serious hard drive failure, and all the data in it was lost. As a result, I was prompted to reinstall my server. The “hydrogen” server was down from  around the 2:00 UTC to around 17:00 UTC (11th September).

Fortunately, I had a backup of the blog contents from 8th September. Reinstalling WordPress and all the software required to run it (Nginx, PHP, MySQL) was still a big hassle, and looks like the PageLines settings aren’t saved in the backup WordPress generated, so I didn’t feel like setting up PageLines again and the blog is back to a white-ish theme (which I have already modified to include a widget area at the bottom…).

It also looks like restoring a backup results in subscribers being sent an email per each post restored… fortunately, the blog only has one subscriber at the moment so this wasn’t a major spamming event.

Oh, and I almost forgot: tnyCloud is gone, but that shouldn’t be a problem since nobody seemed to use it, not even me. Also, since I was forced to move to a smaller VPS, I no longer had enough disk space to host a service like tnyCloud, so basically this hard drive failure saved me the hassle of declaring it dead.

I hope the prolonged downtime didn’t annoy you much. The tny.im URL shortener is hosted on another server which was up all the time – UptimeRobot tells me it has a whooping 99.97% uptime ratio. Now that it already has the three nines, it’s time to work to get the five nines 😉

A server reload is nothing compared to what happened eleven years ago.

Presenting tny.im

Yesterday I wrote a post saying l.f.nu was down… and it still is.
So I bought a domain and moved my URL shortener to yet another domain:

http://tny.im

This time, I’m sure it will be up for at least one year – if it goes down within this period, at least it won’t be because of the domain, as that’s paid for an year already. I didn’t pay it, some friends at Cloudstg did – I’ll pay them back gradually, by advertising their services and such. Again, thanks for investing $11 on my service: if it weren’t you, I’d have to spend my savings on buying this short domain, which would leave me with no money to renew this .com domain next October.

The tny.im domain is as long as l.f.nu, but with less dots, nicer, and since it’s a top level domain and not a subdomain, I have much more control over it. This is a important point, as I plan on adding IPv6 support to tny.im, and a FreeDNS subdomain wouldn’t let me have multiple records on a subdomain. With a real TLD, I can have both an A and an AAAA record for the same domain.

Like it was with the transition from 4.l.to to l.f.nu, no data has been lost, and 4.l.to and l.f.nu links work as long as you change the domain to tny.im. Statistics, link editing, etc. all work.

I hope you enjoy tny.im, and remember, this time it’s for real: the shortener will be around for more than a year, assuming I can get enough profit from it to keep paying for the domain. Having me profiting with tny.im only depends on you – by using my shortening service, you’ll help me earn some cents from ads (but, please, don’t click-bomb them!), which I’ll use to renew the domain and eventually pay for server(s), in order to offer you an even better service.

Again, I hope the ads are not annoying… if they are, make sure to drop me a line so I can fix them.