Also, people underestimate the power of serving out of RAM. It's not unreasonable to serve 20-30K QPS off a single server if the work it needs to do is limited to minimal request parsing and fetching some data from main memory. That's about 2.5 billion requests/day, fully loaded. Granted, I'm thinking something more like memcached than a fully-formed webserver, but an in-memory webserver that stores its data in hashtables (like news.yc) and has a really fast templating languages, or just writes output directly to the socket, could probably come close.
I use redis for this exact reason -- I prerender over 2,000 page templates twice a day, and store them in RAM. The app server has to do a little processing before sending the pages to users -- it picks a different template depending on whether the user's logged in or not, and then substitutes the user's info into the template (for logout/profile links). The session info is also stored in redis. This lets me reboot the server and be ready to serve pages again almost as soon as it's back up. With all the data, redis uses about 300-400MB RAM on a 64bit Debian VM.
I use a VPS for my site, and on a VPS, the only thing you're allocated that you can depend on always being available is RAM. The processor cores might be shared with a busy user, and you can't always depend on high disk I/O speeds.
Why should it be? You can get a long long way by treating the filesystem as a database. The first engineers at Amazon used the same technique a lot, as do I.
If it was always worse then every developer doing this must be stupid. Here are some ways in which a filesystem is "better":
- Zero administration
- Only configuration setting is the directory
- Trivial to test
- Trivial to examine with existing tools, backup, modify etc
- Works with any operating system, language, platform, libraries etc
- Good performance characteristics and well tuned by the operating system
- Easy for any developer to understand
- No dependencies
- Security model is trivial to understand and is a base part of operating system
- Data is not externally accessible
Many existing databases have attributes that aren't desirable. For example they tend to care about data integrity and durability, at the expense of other things (eg increased administration, performance). For a use case like HN, losing 1 out of every 1,000 comments wouldn't be that big a deal - it isn't a bank.
Consider the development, deployment and administrative differences between doing "hello world" with a filesystem versus an existing database. Of course this doesn't always mean filesystems should be used. Developers should be practical and prudent.
Just a nitpick. A "flat file" database suggests encoding all the data to a single file.
Using a files system as a database is a little different as file systems are databases in their own right.
The question to ask is "is the data I want to store in my 'database' enough like the data stored in a filesystem that I can just use the filesystem as my database?"
- Has average latency over 500ms when not under load
- Performs quite poorly under load (I hate to bring it up, but the most recent example was Aaron Swartz's passing. Anyone who used HN then to get news knows how poorly HN performs under load)
- Is restarted every week or two because it leaks memory
- Keeps XSRF tokens in memory and loses them across restarts
- Doesn't have a full markup language
HN is quite poorly-featured compared to typical commenting sites. People use HN because pg is here. He could remove half the features on the site (bold & italics... what features are there even to remove beside nested commenting?) and retain 90% of the audience.
Nothing personal against pg - but I'm here more because of everyone else - the caliber of the discussion, for a news/tech site is quite high, IMHO - and that's due to everyone, not just pg.
Well, I guess we can agree to disagree. HN is is popular for me because of the participants. pg, as epic and central as he is to ycombinator, doesn't play that much of a role on HN in terms of moderating and directing conversations, or even, in recent years, participating that much.
With regards to the commenting site itself, I can think of no more viscerally enjoyable a forum I've ever participated in, with the possible exception of *Forum on MTS. There is nothing whatsoever that I would change about it, with the one possible exception of tweaking the markup so you could add fixed-width text/lists that wrapped over multiple lines. It's the only additional feature I've ever wanted out of HN. There is beauty in it's simplicity. [Edit: Okay, I would also move the upvote/downvote arrows a bit for mobile usage. It's almost impossible to hit the right one without a lot of zooming]
And, with rare exception of a MSM hit, the performance is more than adequate for an environ that should be encouraging reading, digesting, and composing.
You're probably seeing the artificial delay introduced for commenters that can't maintain at least a 4.0 comment point average or people who aren't signed in. Site runs like butter for me.
Just as one example, I think you could make a fairly convincing case that the official version of git uses a "filesystem as a database" system with great success.
I doubt it would be improved by using something more "proper".
Well, no. There are some exceptions, but most databases add a whole lot of bloat you don't necessarily need. Simple files can be just as fast or even faster than using a big database - which is the most important metric to me.
Even if that was true, I'd still tell people to start with flat files for a new project. It's like the advice to do a job yourself before hiring for it: You'll be better equipped to judge how well a database is managing your data if you've already done it yourself.
I find it interesting how easy it is to criticize a functioning system based on some aspect of non-conformance with some hypothetical ideal. The idea that HN wouldn't run on a database is no less astounding than the idea that much of facebook runs on PHP. Design of real systems is often messy and imperfect and deviates from the ideal due to necessity of optimizing one or another factor that may not be obvious.
Keep in mind that a single modern physical server that is decently configured (12-16 cores, 128GB of ram) is the processing equivalent of about 16 EC2 m1.large instances.
