December 2012 Summaries

The system keyspace in Cassandra 1.2 contains tables with information about the cluster and its nodes, including gossip data, tokens for virtual nodes, batch logs, hinted handoffs, index creation status, node IDs, range transfer status, and request traces. These tables correspond to CREATE TABLE syntax and provide drivers with necessary introspection without relying on old Thrift calls. However, users should be cautious when updating data in the system keyspace as schema changes are not allowed.

Hinted Handoff is an optional feature in Cassandra that stores hints locally when a write fails due to unavailable replica nodes. These hints are later replayed to the recovering node, ensuring data consistency and availability. However, hinted handoffs do not count towards ConsistencyLevel requirements and cannot satisfy all types of writes. While it helps maintain performance during failures, it is not a substitute for regular repair operations as hardware failure can still lead to inconsistencies.

Cassandra 1.2 introduces performance optimizations, particularly for clusters with large amounts of data per node. It moves storage engine internal structures off-heap to manage them manually in native memory. Compression metadata and per-row bloom filters are moved off-heap, improving overall performance. Additionally, the update process for indexed columns has been improved, and leveled compaction now supports multiple concurrent compactions across different sstable ranges. The Murmur3Partitioner is introduced as a new default partitioner, offering negligible performance gains over the RandomPartitioner. Streaming optimization allows the kernel to handle data transfers when entire compressed blocks are involved. Hinted handoff in Cassandra 1.2 enables faster recovery of replicas by allowing concurrent delivery of many hints to the target replica.

Cassandra 1.2 introduces new and improved configuration options, including separate timeouts for single-row reads, range scans, writes, truncation, and miscellanea. The update also adds the ability to handle request overload by dropping requests that have timed out before being processed. Additionally, SSL support is extended to client connections, and bloom filters are configured on a per-table basis with different false positive chances for leveled and size-tiered compaction strategies. These changes provide more fine-grained control and improved performance in Cassandra 1.2.

Cassandra 1.2 introduces virtual nodes (vnodes), a new feature that changes the token selection paradigm from one token or range per node to many per node, allowing for smaller and non-contiguous ranges. This results in faster rebuilding of nodes after failure and more efficient disk failure handling due to parallelized validation phases and incremental data transfer. Vnodes also simplify the use of heterogeneous machines in a cluster by enabling proportional allocation of vnodes based on machine capacity, maintaining balance during transitions from older to newer hardware.