company_id,dateA - commencement of working intervaldateB - end of working intervalSo you can point out any lunches of 13:30 and the end of the working day on Friday at 4:45, instead of the usual 5:00.In the planning of data storage, we need to know what future requests will be. Why keep the data? What do they want?Your clarification of how the data will be used will help us to respond more accurately.

I've had jbd2 problems - sometimes I've been so loaded that led to the re-launching of demons with monit. After long agonizing, it became clear that the load was created mysql from many INSERT and UPDATEs. Discharge after each transaction. My base is InnoDB. Writing in mysql construction:innodb_buffer_pool_size = 384M innodb_log_file_size = 16M innodb_flush_log_at_trx_commit = 2 Restart mysql and it all started flying. On the 1G operational server. Base 25m.

This is the maximum amount of memory that will consume one rolled scenario. If this value is exceeded, it will be forced to stop with errorFatal Error: Allowed Memory Size of ... Bytes ExhaustedHowever, it is the maximum allowable volume, the actual memory of the squeaks launched will be determined by their need and will be less than that.

Read memcache (and his clustering), GFS, MYSQL clustering. We read about the organization. livejournal ♪ http://www.livejournal.com/doc/server/ Also, we can draw attention to CDN technology. Read eAccelerator and xCache. ♪ ♪Read the principle of work nginx and apache (apache - backend; nginx - frontend).Well, maybe the most important thing is to drop the HTML code straight.

Used, it's normal.especially when all the application is removed on one server and there is a surplus of operational memory. But it's less convenient if you have a system that grows up and has to scale it. As long as it's small, you don't have enough users and there's little information in the sessions. However, if the system grows up, it will be difficult for you to operate a local hard drive in your memory.(1) Disk is on the same night, if you have 10 more, how will you access the sessions? Even if you arrange an exchange, you'll have a mass of internal traffic between the nodes, and the more you bring them, the traffic will be higher, up until you've blocked them. It is therefore best to develop a system with a scale-up focus.(2) Let's say the size's up so much that you don't have the data on the same server, there's no memory of how you're gonna scale this system? How fast can this disk be spread to two servers? There's no problem with NoSQL-Server-- they're clustering out of the box, and what do we do here? Write your own server?When the system's big, I don't want the original solutions, I want to be reliable, because the different problems are so many and I want to at least be deleted as a class. Disclosure in operational memory is a good budgetary decision if you don't plan to grow out of one server. If this is possible, it would be better to immediately introduce the layers of the application (database, nosql, annex, front-server, statistic server, kash, line of assignments, etc.) and to place them on a home, it would be easier to transfer the system to several servers.If the system does not grow, if you can predict the workload for years ahead, the disc in the operational memory of the session is a perfectly normal solution.

Resources of any end server. Every connection to the client consumes memory, processor, part of the network. So if you have a large number of users and connections, you try to use a few servers by assigning a load between them.The first thing that comes to mind is to separate static files (CSS, JavaScript), annex and database. We put them on different servers.Further, we increase the grouping of servers for static files by assigning a separate dose to them, either a balancer can be placed in front of them or each server has its own sub-distance, choosing them by accident. If the pressure on the statistic is increasing, you're just increasing the number of servers in the statistic pool.Next, we'll scale up the application server, and we'll increase their grouping. You can't afford to keep the user's media files on the same server, otherwise you're gonna have to use the avatar on one server to spread it to all the servers of your bullet. If they're hundreds, you'll forget all the channels. That's why all the media files are stored on individual servers, better in some infinite vault (Swift, S3, Ceph). In this case, all the servers with your application will only have the code and their number can be increased by comparing the balancer load.Next, the database is the hardest place. Usually, any database has a replication mechanism when you connect databases to the chain, recording data in one server(s) on the replication mechanism, it is reproduced on all sub-servers (slaves). There are two problems: the information is reproduced with some delay on the one hand, on the other, so you only scale reading, the record is not large. If you have a small number of users who write, you'll still be able to live with one master if their millions are the ones that should be scaled differently. There are several mechanisms: a replication by type of ring, when you combine a number of masters in the ring, say 10, and every tenth id-nick writes to your individual server, which is divided by the rest of the replicative masters (not seen in the living mechanism - very hemorrhoid for many reasons). We can calculate the user's hex, for example, by his name and distribute users to different servers, for example, you have 256 database servers, we're computing md5-hash, we're taking the first two hexa symbols, they're about to be 256 servers. All users with suitable hesh can be serviced by a dedicated database server. Usually, they try to select a mechanism that distributes users more evenly.Next thing you know, you're starting to set up separate software application modules, advertising runs your pool of servers, commentaries, your own, your news, your counters. It's only in the past that you don't want to do this unless you need real-time enumerators you need a separate system for them. Just spend time and effort. However, design the application so that it has a database, an application server, a statistician and a media file can be serviced by different servers, and better server pools are good.

React.php tries to process incoming enquiries asynchronously using non-locking input/drip. This means that until one request is fully processed (and the response is given to the client) the process can be processed start. processing a few more requests. In general, React.php is very similar to node.js.If you use the second method (leasing the field of the general context), then it might end very well. For example, you will not be able to agree on the headings of the request/response if this process is in some context.The only right in this situation would be to create a separate context for each request and to kill it when the processing of the request is completed and the response is given to the client.

For example, there is a request for OBD, its results can be retained in Redis in the form of JSON, and the following applications will already take it from Redis and, where necessary, make the cash disable. - How right is that?100 per cent, but the timely handicap of the kash is a pretty big shuttle.Is it worth keeping data in JSON?In fact, I'm a little bit of a Corobiter to keep this line, but there's no specific choice, and it doesn't have to affect anything at all.We sometimes need a page encrypted not only in JSON, but a piece or a whole HTML. - Is it right to keep these pieces in Redis?No, you have a SPA, all the rendering should be on the client.There's an idea to implement a session-storage in Redis, how to do it right Is that right or not? How will it be better?If your session is considered to be a permanent repository, it's not right because Redis is not meant to keep the data permanently (whether he can turn the data on the CD from time to time). Instead, it is better to organize a two-tier (or even three-level) repository from Redis-OBD (In-Memory - Redis-OBD in the case of tri-level interconnections).In which case should the cash be recorded in the CD files?Normally, there's a database. There's a way to throw the results of the heavy requests so that all their legs can be seen, but even in this case, I'd organize an interim Redis-Slave.Shouldn't we be drinking what's described in redis?The more you hang in the cash, the better, but you watch the number of melted operational memory from time to time.Maybe there's some guidance on working with high-load sockets or something on the basis of such a project.Remember two things.The project has to be scaled up horizontally, i.e. simple addition of new servers. It follows that changes on one server should be visible to all others (this is achieved through the use of one Redis and OBD)There's a thing like dogpile effect, which means a bunch of users who at the same time request the same resource. In the case of the three-level repository described above, 100 simultaneous users are not detected in the cash and are laminated (all 100) to the OBD. In the case of 20 simultaneous connections in the OBD system and each request takes 30ms, the last user will receive data not before 150ms (which ideally is, in fact, there would be a considerable amount of data.Oh.That's not good.

http://nginx.org/ru/docs/http/ngx_http_proxy_module.html#proxy_cache right robots.txt, and http://nginx.org/ru/docs/http/ngx_http_limit_conn_module.html ip.UpdatingThe hanging of dynamically emerging pages will reduce the number of requests to bequenda, thus freeing systemic resources. ♪ robots.txt increase time through which the Kraulers enter the site. With help limit_conn_module The number of simultaneous compounds with Google ip addresses can be limited and the burden on the server can be reduced.

IMOW. NoSQL won't come from the word "real." They're trapped in a different, more "variable" type. Under your simple structure, you need a relay database, because data from a large number of sources are that server: MSSQL, MySQL, Oracl - most likely any of them will.

A serious question. So now you have an average of 6.5 SMSs. Probably about 60 at the peak.Accordingly, the new system can be expected at a peak of about 10,000 messages per second. IMHO is quite a lot even from the network, and how many queries to the base are there? First of all, you need to model this load and see if, in principle, the existing solution or design something new with a distributed base and several physically different reception points.Read different materials on the subject. https://ru.wikipedia.org/wiki/%D0%9F%D1%80%D0%BE%D0%B1%D0%BB%D0%B5%D0%BC%D0%B0_10000_%D1%81%D0%BE%D0%B5%D0%B4%D0%B8%D0%BD%D0%B5%D0%BD%D0%B8%D0%B9 ♪ Well, I guess. https://ru.wikipedia.org/wiki/%D0%92%D1%8B%D1%81%D0%BE%D0%BA%D0%B0%D1%8F_%D0%B4%D0%BE%D1%81%D1%82%D1%83%D0%BF%D0%BD%D0%BE%D1%81%D1%82%D1%8C (this reference is just a term symbol, really needs to read others Technical The materials (I hope you need to googly) should not be overlooked. Just take into account, there are no cheap solutions to such tasks.(sighs)It's just that I haven't been able to change the comments (and with the retention of the references) in response, so I left one, or that's not clear, because I can't remove other people's comments.

Cop towards Linux in Nginx+PHP+MySQL+Memcache. I'm sure that the same server will help with the highest number of users.

Anyway, I've been struggling with this six months and I've identified recommendations for optimizing wordpress:(1) Add the coke(2) Optimize schedule (Image Magic)(3) Insert Gzip(4) Move js in footer and make asynchronous loading(5) Put the main CSS in the header negatively, the rest of us load asynchronously.Examples. Wordpress https://developers.google.com/speed/pagespeed/insights/?hl=ru&url=https%3A%2F%2Fwob.su%2Fblog%2F&tab=desktop SantaFox https://developers.google.com/speed/pagespeed/insights/?hl=ru&url=https%3A%2F%2Fwob.su%2F It's not perfect, it's still working, but there's a green line.

If such a problem is possible, how is it handled?Paul AUTO_INCREMENT / IDENTITY in the OBD (depending on driving).The example itself is the typical multi-point task, if the mechanism is to be implemented, it is necessary to implement it. lock As long as someone uses it, everyone else is waiting. As I recall, PHP can use the file reader rooms - while one process put the elbow on the file, the others are waiting. I understand that there is no normal way to deal with this problem in PHP.

If the volumes are so large, it makes sense to split the transferred data into smaller parts. Or the second option: the first request to the main base is to receive a list of id new lines in it, and then to ask the principal only ones that are not in the current database.

If there's no desire to spend money on services, you can do that with a jmeter. Examples include: http://habrahabr.ru/post/140310/ http://habrahabr.ru/post/84190/

Just a little bit ofcorn, you can find an article. http://qnatech.wordpress.com/2010/03/05/how-much-traffic-can-my-web-application-handle/ It assumes that the default volume has 200 workflows. And I think that 0.2 seconds will be the normal processing time of one request.From this point of view, they receive between 200 and 1,000 requests per second.As it is planned to make a request once a minute, 200*60 = 12,000 to 60000 on-line.Since not all the lasers are used continuously, so more users can actually be serviced, and 5 to 7 per cent of users are generally allowed to be online. It's still 20 times the collovo, from 240000 to 1200000. But I think you'll be on the CD.You can come on the other side. Let one mark be kilobite. On average, the user will keep a hundred notes (some more, less). Total 100 kilobite on the user (the profile and overhead costs are negligible). 30 Gigabite for 0.1 Mbeit - 300,000 users (and it seems that java itself needs to be installed and logs stored. ♪ ♪ )Therefore, in the current iron, I would not count on a user ring over 50 to 100,000.

load average is not just a lack of processor. LA raises all processes that are waiting for something, such as input-outs, or a response from the deleted server.In fact, the LA itself doesn't say much, it's interesting to change it over time.

Option: https://ru.wikipedia.org/wiki/Pluggable_Authentication_Modules and him. http://linux.die.net/man/5/limits.conf

Intel(R) Xeon(R) CPU X5650 @ 2.67GHz24GbHe's got a lot of Haproxy, a lot of queries, an average LA ~2-3, I don't complain. Ugh.