General

Configuration file is parsed by DatabaseDescriptor (which also has all the default values, if any)
Thrift generates an API interface in Cassandra.java; the implementation is CassandraServer, and CassandraDaemon ties it together (mostly: handling commitlog replay, and setting up the Thrift plumbing)
CassandraServer turns thrift requests into the internal equivalents, then StorageProxy does the actual work, then CassandraServer turns the results back into thrift again
- CQL requests are compiled and executed through QueryProcessor. Note that as of 1.2 we still support both the old cql2 dialect and the cql3, in different packages.
StorageService is kind of the internal counterpart to CassandraDaemon. It handles turning raw gossip into the right internal state and dealing with ring changes, i.e., transferring data to new replicas. TokenMetadata tracks which nodes own what arcs of the ring. Starting in 1.2, each node may have multiple Tokens.
AbstractReplicationStrategy controls what nodes get secondary, tertiary, etc. replicas of each key range. Primary replica is always determined by the token ring (in TokenMetadata) but you can do a lot of variation with the others. SimpleStrategy just puts replicas on the next N-1 nodes in the ring. NetworkTopologyStrategy allows the user to define how many replicas to place in each datacenter, and then takes rack locality into account for each DC -- we want to avoid multiple replicas on the same rack, if possible.
MessagingService handles connection pooling and running internal commands on the appropriate stage (basically, a threaded executorservice). Stages are set up in StageManager; currently there are read, write, and stream stages. (Streaming is for when one node copies large sections of its SSTables to another, for bootstrap or relocation on the ring.) The internal commands are defined in StorageService; look for registerVerbHandlers.
Configuration for the node (administrative stuff, such as which directories to store data in, as well as global configuration, such as which global partitioner to use) is held by DatabaseDescriptor. Per-KS, per-CF, and per-Column metadata are all stored as parts of the Schema: !KSMetadata, !CFMetadata, ColumnDefinition. See also ConfigurationNotes.

Some historial baggage

Some classes have misleading names, notably ColumnFamily (which represents a single row, not a table of data) and !Table (which represents a keyspace).

Write path

StorageProxy gets the nodes responsible for replicas of the keys from the ReplicationStrategy, then sends RowMutation messages to them.
- If nodes are changing position on the ring, "pending ranges" are associated with their destinations in TokenMetadata and these are also written to.
- ConsistencyLevel determines how many replies to wait for. See WriteResponseHandler.determineBlockFor. Interaction with pending ranges is a bit tricky; see https://issues.apache.org/jira/browse/CASSANDRA-833
- If the FailureDetector says that we don't have enough nodes alive to satisfy the ConsistencyLevel, we fail the request with UnavailableException
- If the FD gives us the okay but writes time out anyway because of a failure after the request is sent or because of an overload scenario, StorageProxy will write a "hint" locally to replay the write when the replica(s) timing out recover. This is called HintedHandoff. Note that HH does not prevent inconsistency entirely; either unclean shutdown or hardware failure can prevent the coordinating node from writing or replaying the hint. ArchitectureAntiEntropy is responsible for restoring consistency more completely.
- Cross-datacenter writes are not sent directly to each replica; instead, they are sent to a single replica, with a Header in MessageOut telling that replica to forward to the other ones in that datacenter
on the destination node, RowMutationVerbHandler uses Table.Apply to hand the write first to the CommitLog, then to the Memtable for the appropriate ColumnFamily.
When a Memtable is full, it gets sorted and written out as an SSTable asynchronously by ColumnFamilyStore.maybeSwitchMemtable (so named because multiple concurrent calls to it will only flush once)
- "Fullness" is monitored by MeteredFlusher; the goal is to flush quickly enough that we don't OOM as new writes arrive while we still have to hang on to the memory of the old memtable during flush
- When enough SSTables exist, they are merged by CompactionManager.doCompaction
  - Making this concurrency-safe without blocking writes or reads while we remove the old SSTables from the list and add the new one is tricky. We perform manual reference counting on sstables during reads so that we know when they are safe to remove, e.g., ColumnFamilyStore.getSSTablesForKey.
  - Multiple CompactionStrategies exist. The original, SizeTieredCompactionStrategy, combines sstables that are similar in size. This can result is a lot of wasted space in overwrite-intensive workloads. LeveledCompactionStrategy provides stricter guarantees at the price of more compaction i/o; see http://www.datastax.com/dev/blog/leveled-compaction-in-apache-cassandra and http://www.datastax.com/dev/blog/when-to-use-leveled-compaction
See ArchitectureSSTable and ArchitectureCommitLog for more details

Read path

StorageProxy gets the endpoints (nodes) responsible for replicas of the keys from the ReplicationStrategy as a function of the row key (the key of the row being read)
- This may be a SliceFromReadCommand, a SliceByNamesReadCommand, or a RangeSliceCommand, depending on the query type. Secondary index queries are also part of RangeSliceCommand.
StorageProxy filters the endpoints to contain only those that are currently up/alive
StorageProxy then sorts, by asking the endpoint snitch, the responsible nodes by "proximity".
- The definition of "proximity" is up to the endpoint snitch
  - With a SimpleSnitch, proximity directly corresponds to proximity on the token ring.
  - With implementations based on AbstractNetworkTopologySnitch (such as PropertyFileSnitch), endpoints that are in the same rack are always considered "closer" than those that are not. Failing that, endpoints in the same data center are always considered "closer" than those that are not.
  - The DynamicSnitch, typically enabled in the configuration, wraps whatever underlying snitch (such as SimpleSnitch and PropertyFileSnitch) so as to dynamically adjust the perceived "closeness" of endpoints based on their recent performance. This is an effort to try to avoid routing more traffic to endpoints that are slow to respond.
StorageProxy then arranges for messages to be sent to nodes as required:
- The closest node (as determined by proximity sorting as described above) will be sent a command to perform an actual data read (i.e., return data to the co-ordinating node).
- As required by consistency level, additional nodes may be sent digest commands, asking them to perform the read locally but send back the digest only.
  - For example, at replication factor 3 a read at consistency level QUORUM would require one digest read in additional to the data read sent to the closest node. (See ReadCallback, instantiated by StorageProxy)
- If read repair is enabled (probabilistically if read repair chance is somewhere between 0% and 100%), remaining nodes responsible for the row will be sent messages to compute the digest of the response. (Again, see ReadCallback, instantiated by StorageProxy)
On the data node, ReadVerbHandler gets the data from CFS.getColumnFamily, CFS.getRangeSlice, or CFS.search for single-row reads, seq scans, and index scans, respectively, and sends it back as a ReadResponse
- The row is located by doing a binary search on the index in SSTableReader.getPosition
- For single-row requests, we use a QueryFilter subclass to pick the data from the Memtable and SSTables that we are looking for. The Memtable read is straightforward. The SSTable read is a little different depending on which kind of request it is:
  - If we are reading a slice of columns, we use the row-level column index to find where to start reading, and deserialize block-at-a-time (where "block" is the group of columns covered by a single index entry) so we can handle the "reversed" case without reading vast amounts into memory
  - If we are reading a group of columns by name, we still use the column index to locate each column, but first we check the row-level bloom filter to see if we need to do anything at all
- The column readers provide an Iterator interface, so the filter can easily stop when it's done, without reading more columns than necessary
  - Since we need to potentially merge columns from multiple SSTable versions, the reader iterators are combined through a ReducingIterator, which takes an iterator of uncombined columns as input, and yields combined versions as output
  - Single-row reads use CollationController to determine which sstables are relevant -- for instance, if we're requesting column X, and we've read a value for X from sstable A at time T1, then any sstables whose maximum timestamp is less than T1 can be ignored.

In addition:

At any point if a message is destined for the local node, the appropriate piece of work (data read or digest read) is directly submitted to the appropriate local stage (see StageManager) rather than going through messaging over the network.
The fact that a data read is only submitted to the closest replica is intended as an optimization to avoid sending excessive amounts of data over the network. A digest read will take the full cost of a read internally on the node (CPU and in particular disk), but will avoid taxing the network.

Deletes

See DistributedDeletes

Gossip

based on "Efficient reconciliation and flow control for anti-entropy protocols:" http://www.cs.cornell.edu/home/rvr/papers/flowgossip.pdf
See ArchitectureGossip for more details

Failure detection

based on "The Phi accrual failure detector:" http://vsedach.googlepages.com/HDY04.pdf

ArchitectureInternals

General

Some historial baggage

Write path

Read path

Deletes

Gossip

Failure detection

Further reading