ZooKeeper maintains a order when processing requests:
We will explain the three aspects of ZooKeeperServer: request processing, data structure maintenance, and session tracking.
If the request is just a query, it will be processed by ZooKeeper and returned. Otherwise, the request will be validated and a transaction will be generated and logged. This the request will then wait until the request has been logged before continuing processing.
Requests are logged as a group. Transactions are queued up and the SyncThread will process them at predefined intervals. (Currently 20ms) The SyncThread interacts with ZooKeeperServer the txnQueue. Transactions are added to the txnQueue of SyncThread via queueItem. When the transaction has been synced to disk, its callback will be invoked which will cause the request processing to be completed.
We guarantee that changes to nodes are stored to non-volatile media before responding to a client. We do this quickly by writing changes as a sequence of transactions in a log file. Even though we flush transactions as a group, we need to avoid seeks as much as possible. Also, since the server can fail at any point, we need to be careful of partial records.
We address the above problems by
As the server runs, the log file will grow quite large. To avoid long startup times we periodically take a snapshot of the tree of DataNodes. We cannot take the snapshot synchronously as the data takes a while to write out, so instead we asynchronously write out the tree. This means that we end up with a "corrupt" snapshot of the data tree. More formally if we define T to be the real snapshot of the tree at the time we begin taking the snapshot and l as the sequence of transactions that are applied to the tree between the time the snapshot begins and the time the snapshot completes, we write to disk T+l' where l' is a subset of the transactions in l. While we do not have a way of figuring out which transactions make up l', it doesn't really matter. T+l'+l = T+l since the transactions we log are idempotent (applying the transaction multiple times has the same result as applying the transaction once). So when we restore the snapshot we also play all transactions in the log that occur after the snapshot was begun. We can easily figure out where to start the replay because we start a new logfile when we start a snapshot. Both the snapshot file and log file have a numeric suffix that represent the transaction id that created the respective files.