How to handle database connections?

How do database connections get allocated to different services in LinkedIn and Twitter?

• From what I know, both companies have many backend services running on their system, communicating with each other and with front-end. Many of these services would be relying on same database which produces resource contention due to limited number of simultaneous connections available. So how is it decided which service gets what number of connections or does there exist a completely different methodology to share db resource among services? PS: I am particularly interested in LinkedIn and Twitter because they are quite open about their engineering efforts.

• Answer:

  A good example at Twitter is our use of https://github.com/twitter/gizzard with http://twitter.github.io/finagle/. The simple answer is services don't talk directly to databases. Between the DB and services that talk to the database is another service (a Gizzard app).  Services speak thrift RPC to this app.  The Gizzard app talks SQL to the backend databases and returns the result. This gives us many things (like with Gizzard we have a sharding and consistency framework).  But in the context of managing DB questions, Finagle allows us to cap the number of connections and internally queue requests when the limit is reached.  We can also embed other business logic in this app layer to do things like priority queuing for different types of operations.