The load balancer, combined with the performance monitor, directs requests to servers. This is necessary because when we have multiple servers, we want to make sure we're using each one as fully as possible. Thus, it's useful to have a component that, for example, prefers to send jobs to servers with empty queues rather than servers with other requests queued up.
We need at least one web server in order to be able to generate responses to requests. This may involve loading data from the database, storing data in the database, performing a calculation, or pretty much anything a server can do. We want multiple servers so that our web site has low latency even given lots of simultaneous requests, so that we can efficiently process requests concurrently.
We want a database because files stored in a normal filesystem aren't organized as optimally as they could be. Database allow us to optimize for specific patterns of reading/writing (e.g. tables or indexes in SQL databases). The database has multiple slaves and one master so that multiple reads (from slave DBs) can occur concurrently, while writes have to be processed by the master DB (and propagated to slaves). This allows reads (which are much more common) to occur efficiently, while still ensuring eventual consistency.
Finally, we have memcached servers because even with a highly optimized database, reading from a DB is generally the slowest part of generating a response. Thus, we want to cache the most frequently accessed data in memcached to ensure low latency. We can also use memcached to cache pretty much anything, such as expensive computations or page renderings.
This comment was marked helpful 5 times.
nrchu
Note that this particular DB implementation makes use of replication of data. This is particularly useful because most applications have many more reads than writes and we also are okay with relaxed consistency on read/write ordering. For use cases where writes are about as common as reads, or space might be more limited, partitioning the database is viable as well.
The load balancer, combined with the performance monitor, directs requests to servers. This is necessary because when we have multiple servers, we want to make sure we're using each one as fully as possible. Thus, it's useful to have a component that, for example, prefers to send jobs to servers with empty queues rather than servers with other requests queued up.
We need at least one web server in order to be able to generate responses to requests. This may involve loading data from the database, storing data in the database, performing a calculation, or pretty much anything a server can do. We want multiple servers so that our web site has low latency even given lots of simultaneous requests, so that we can efficiently process requests concurrently.
We want a database because files stored in a normal filesystem aren't organized as optimally as they could be. Database allow us to optimize for specific patterns of reading/writing (e.g. tables or indexes in SQL databases). The database has multiple slaves and one master so that multiple reads (from slave DBs) can occur concurrently, while writes have to be processed by the master DB (and propagated to slaves). This allows reads (which are much more common) to occur efficiently, while still ensuring eventual consistency.
Finally, we have memcached servers because even with a highly optimized database, reading from a DB is generally the slowest part of generating a response. Thus, we want to cache the most frequently accessed data in memcached to ensure low latency. We can also use memcached to cache pretty much anything, such as expensive computations or page renderings.
This comment was marked helpful 5 times.
Note that this particular DB implementation makes use of replication of data. This is particularly useful because most applications have many more reads than writes and we also are okay with relaxed consistency on read/write ordering. For use cases where writes are about as common as reads, or space might be more limited, partitioning the database is viable as well.
This comment was marked helpful 0 times.