6 ms·

Ask HN: Server under heavy load. Any downside to this hack?

8 points by webvet 13y ago

The server hosting one of our live apps is under high loads and gets into a state where it stops responding to http requests several times a day. These outages last typically anywhere from a few minutes all the way up to half an hour or so. The major culprit here is MySQL load and we've been working on optimizing this (those interested can refer to my previous thread on this: https://news.ycombinator.com/item?id=6348903). For now we've found that restarting the httpd and mysqld services brings things back to normal almost immediately. While we continue to work on a - more elegant - solution, we're thinking of writing a bash/shell script that runs on hourly cron, checks load average (via uptime or /proc/loadavg) and if found higher than a threshold, restarts the services. Can anyone think of any downside to this (used at least as a temporary measure)?

joshbaptiste 13y ago
Crude check but since it's temporary, an hour seems far too long an interval, you check at 10GMT , by 10:05GMT your server is in trouble and has 55 minutes to crap out. I would check every */5 minutes at least.
- webvet 13y ago
  Good call - thanks :)
cbhl 13y ago
I would be really worried about losing requests in-flight or that take a long time to run. Is it prohibitively expensive/time consuming to get (or borrow) a bigger machine (on EC2, or in your colo, or what have you) to run MySQL on until you've figured out how to shard / scale out your application?
- webvet 13y ago
  We are actually more than worried too, because we _know_ we're losing plenty of requests daily so trying to cut down the losses temporarily while we figure out a more long-term solution. For the rest, this thread might give you some more info on the issue: https://news.ycombinator.com/item?id=6348903 https://news.ycombinator.com/item?id=6348903
lutusp 13y ago
> For now we've found that restarting the httpd and mysqld services brings things back to normal almost immediately. You need to examine the restart process and analyze why it resolves the issue. If the reason is the abandonment of dead parasitic processes and memory leaks, you need to find out why and correct them. If the reason is that the restart unceremoniously drops all the current transactions, you need to increase capacity. > Can anyone think of any downside to this (used at least as a temporary measure)? I certainly can -- a bunch of really irritated visitors, whose transactions are abandoned. But that's only true if that is actually what's going on. Make sure you don't have software issues that are preventing efficient operation. If that's not the issue, you need to grow with your customer base -- increase server capacity.
- webvet 13y ago
  >You need to examine the restart process and analyze Like I said in OP, we've identified MySQL to be the primary bottleneck and are already working on resolving this. >whose transactions are abandoned. At least some saving grace here... we're not selling anything... it is purely a content site Thanks for your input.
  - lutusp 13y ago
    >Like I said in OP, we've identified MySQL to be the primary bottleneck and are already working on resolving this. Ah, yes. I remember from your prior post that you have very large databases and table sizes and are considering (or have begun) partitioning the largest tables. It turns out there is an innate partitioning scheme built into the most recent MySQL versions, but it has to be compiled into the running binary by way of a compiler flag: http://dev.mysql.com/doc/refman/5.1/en/partitioning.html http://dev.mysql.com/doc/refman/5.1/en/partitioning.html I apologize if you've already heard this, or if it's already in place as a partial solution.
    - webvet 13y ago
      Yes, we did investigate MySQL's internal partitioning option briefly before deciding to roll our own scheme (which, after working quite well initially, is now beginning to create problems of its own). Perhaps, it is time for us to revisit this. Thanks again Paul.
AznHisoka 13y ago
I've had load issues in Postgres, not MySQL before and it was due to autovacuum running on tables getting updated/inserted frequently. I'm not sure what the equivalent in MySQL is, but if you have a ton of insert/update queries, consider archiving your tables after a certain period of time, so that your main table doesn't have ton of rows. You can consider sharding of course, but also consider sharding the actual tables in the same database. An insert query on a table with 1000 rows will take much much less time than an insert query on a table with 100 million rows, all things considered. Also, consider creating a buffer in the application layer that buffers inserts/updates and executes them once as a single transaction, if they don't need to be executed immediately. It puts less stress on the database. Of course, this would require a lot of rewriting in your app, so not sure if you want to go through this route. Indices are another area. I'm sure plenty of people have told you to optimize your indices, but also consider REMOVING unnecessary indices. Do you have an index on a text column, or multiple varchar columns? Those can be killer after awhile because inserts will slow down. Consider changing indices on varchar columns to indices on an int column by hashing those strings. A quick suggestion: Install NewRelic (it's free for a certain period), and check out the database transactions that are taking up the most CPU load. Sometimes there's that 1 query you overlooked that is table scanning and could be the main culprit. Also, are you using Rails by any chance? If so, there are other areas I can suggest. And please post your server specs. Maybe your VPS just does suck (no offense), and the easiest route is just to upgrade your server.
- webvet 13y ago
  Thanks for your in-depth reply and suggestions. MySQL performance tuning and table structure/query optimization-wise, we've left few stones unturned (but I'm sure there still is a scope for further improvement - isn't there always?). Not using Rails. Server specs: 4 cores, 3.2GB RAM >and the easiest route is just to upgrade your server. Please refer to the other thread for more on this. Thanks again.
  - staunch 13y ago
    Your server barely has more RAM than my phone. Your disk almost certainly has less reliable performance than my laptop's. If you're doing something that makes any money you should get a real server before wasting days of effort or intentionally causing repeated outages.
cpncrunch 13y ago
Restarting the processes isn't really solving the problem. What is the actual bottleneck? It should be possible to figure this out. I think your first problem is that you are using a VPS. You should never use a VPS in a high load situation like this - buy a dedicated server! They only cost about $70/month, which you should be able to afford if you have a successful site. Ideally you should get as much RAM as you can afford and/or an SSD drive. I know you said you didn't want to throw hardware at the problem, but there are limits - you can't run a massive database on crappy hardware and expect it to work smoothly.
- webvet 13y ago
  >What is the actual bottleneck? As mentioned in OP, it is MySQL. >you can't run a massive database on crappy hardware and expect it to work smoothly. Absolutely... but we do wish to milk the VPS for every cent of its worth and we're not sure we've got there yet. Thanks for your input.
  - stevekemp 13y ago
    > As mentioned in OP, it is MySQL. That's not specific enough. Yes, the problem is manifesting itself via MySQL, but the bottleneck will be somewhere else; whether RAM, CPU, or I/O-wait.
    - cpncrunch 13y ago
      Exactly. If the OP can post the output of top, vmstat and iostat both when the problem occurs and after restarting the processes, that would be a start. There might be some simple thing you can tweak to improve things.
      - webvet 13y ago
        Will try and post within a few hours. Thanks.
    - webvet 13y ago
      For now, we're making our code/queries/optimization take the blame rather than the hardware.
trevelyan 13y ago
If you haven't already done this, when the mysql server is having trouble make sure you connect through the terminal and try: > SHOW PROCESSLIST This will show all active queries and the time they have taken to execute. The fact that the server seems to churn to a halt and then work its way through the problem suggests the issues are related to specific queries you can catch this way. Then use the EXPLAIN command on the slow queries to figure out why they are hanging your server and add indexes or tweak that part of your code (avoid joins on large tables, etc.) as necessary.
- webvet 13y ago
  Yes, as mentioned in the other thread, we're way ahead of this - but thanks for your input all the same.
  - trevelyan 13y ago
    k, good luck!
staunch 13y ago
It seems pretty clear that you don't have enough experience on your own to resolve this properly, so call in some help. It may be possible to make an architectural change that significantly reduces the resources you need, or perhaps you'll find that you unavoidably need more resources to do what you want. Someone who knows how to diagnose and analyze this properly can tell you that.
- webvet 13y ago
  >so call in some help. As is being discussed in the other thread.
motilevy 13y ago
while the root cause should be fixed ( sound like you're working on it ). consider using monit for the restarts instead of cron : http://mmonit.com/monit/ http://mmonit.com/monit/
- webvet 13y ago
  Wow!!! Didn't know about this - took a quick look and found it very interesting. Thanks.

cpncrunch 13y ago
I can't really see how your code or queries could cause mysql or apache to gradually slow down and then speed up when you kill the processes. It sounds more like virtual memory of some process is growing too high causing thrashing, or apache has too many children, or you're running out of some resource, or something is leaking memory. Regarding apache/mysql: are you using stock distributions, or did you build them yourself? Are you using any unusual configuration in either of them? You say you kill both mysql and apache, but what happens if you just kill one or the other? I'm just wondering if it is some buggy cgi script that is leaking resources.
- webvet 13y ago
  >I can't really see how your code or queries could cause mysql or apache to gradually slow down and then speed up... Actually, we've ascertained that it is not the code or queries that cause this per se, but it is the traffic patterns - human visitors + SE crawlers... at times, there are (natural) traffic bursts and that's when the server starts choking. apache/mysql - stock distros, with some config tweaks, mainly in my.cnf. >what happens if you just kill one or the other? Good call. Actually, restarting just apache does the trick most of the times. What we suspect is, abrupt flooding in http requests causes MySQL to slow down which causes http requests to get queued up and finally choke the service.
  - AznHisoka 13y ago
    Are you sure you have a database connection pool in place? Ensure the # of active connections remain constant no matter how many http requests are coming in.
  - cpncrunch 13y ago
    Ok, I still think you probably need to spend some money on better hardware, but here are a few ideas for eeking out some more performance from your existing setup: - Limit the max apache child processes/threads to what your server can easily handle. That way during busy periods people will get an occasional error connecting to your server (or wait a while), but at least mysql won't be thrashing your server so it might improve overall performance during the busy times. - Figure out which queries are using the most cpu time and optimize those. - Have a look at your most frequently accessed pages and see if you can optimize database access in those pages. However if your users are still sometimes getting error messages accessing your site you have to ask yourself whether that is going to hurt your business.