This document explains how Ledger works by following the mechanics of a single Put operation.
The Putting of a value happens in 2 major stages:
The client facing API of Ledger is defined in ledger.fidl. There are multiple methods to do put operations, with slightly different semantics. For this example we can use the FIDL method Put(array<uint8> key, array<uint8> value) => (Status status);
.
This FIDL protocol is implemented by PageImpl, which takes a generated callback type (PutCallback) used to pass the completion status back to the client.
The first step to handling the FIDL call is storing the value (which can be large, e.g. multiple GB) onto the disk. The value (which can be passed as a byte array, a VMO, or a socket) is wrapped in a DataSource. The data in the DataSource is cut in chunks (see dynamic chunking for an explanation of the idea, and the splitting implementation for details). The chunks not already on disks are then written to disk.
Clients can register PageWatchers to observe changes to Pages. Before creating the commit, we drain the pending updates to the PageWatchers registered for the given Page connection. After draining the pending updates, we stop the BranchTracker from sending new updates until we've handled the change.
If the Put is part of a transaction, we wait for all PageWatchers to have handled the updates before continuing to the next step. The PageWatchers say that they've handled the OnChange by calling the (generated) OnChangeCallback.
All changes are first written in a Journal before being committed. This includes Put, Delete or Clear operations. Journals can be used to store transactions in progress, in which case the journal is called explicit, or to prepare a single update operation outside of a transaction (implicit journal).
Changes in journals are guaranteed to be seen together: a PageSnapshot and/or a PageWatcher will either contain/see all the changes, or none. If for some reason Ledger is unexpectedly closed, the updates added in any non completed journal are dropped.
When a transaction is completed, a commit is appended to one of the Page's heads.
The BranchTracker of the Page connection is informed of this new commit as to make sure it tracks that branch and not some other branch. Not doing so would make changes seem to “disappear” as clients would see a different branch than the one with their change. The BranchTracker then notifies the relevant registered Watchers of the change. If there are multiple heads, the MergeResolver starts the process of merging the heads.
TODO