Cassandra Admin flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

Murmur3Partitioner

Default partitioner in modern Cassandra. It hashes the partition key with Murmur3 into a signed 64-bit token, and that token decides which nodes own the row on the ring.

Gossip protocol

Peer-to-peer protocol nodes use roughly once per second to exchange heartbeat, status, schema, and load information. A phi-accrual failure detector then decides whether a peer is down.

Virtual nodes

Multiple token ranges assigned to one node. num_tokens sets the count, 16 by default in Cassandra 4.0, giving finer ownership granularity but more streaming sessions during bootstrap.

Coordinator node

The node a client connects to. It computes token ownership, forwards the request to the replica set, and merges responses until the requested consistency level is satisfied.

endpoint_snitch

cassandra.yaml setting naming the snitch class. GossipingPropertyFileSnitch reads datacenter and rack from cassandra-rackdc.properties and spreads that topology to peers through gossip.

allocate_tokens_for_local_replication_factor

Cassandra 4.0 setting that derives evenly distributed tokens for a joining node from a keyspace replication factor, replacing manual token arithmetic and reducing post-bootstrap imbalance.

Seed nodes

Addresses listed in cassandra.yaml that a starting node contacts to learn cluster membership through gossip. They are bootstrap contact points, not a runtime single point of failure.

JVM heap sizing

Set -Xms equal to -Xmx and keep the heap near or below 31 GB to preserve compressed oops. Larger heaps lengthen garbage collection pauses without improving throughput.

Partition key

The first component of the primary key. Its hash selects replica nodes, so all rows sharing a partition key reside together and are fetched in a single partition read.

Clustering columns

Columns following the partition key in the primary key. They sort rows inside a partition and enable range, slice, and ORDER BY queries restricted to one partition.

Static column

A column declared STATIC stores one value shared by every row in a partition, avoiding repeated storage of attributes that are constant across a wide partition.

Secondary index

A per-node index on a non-key column. It suits low-cardinality predicates because every write updates the index on each replica and reads may scatter across all nodes.

NetworkTopologyStrategy

Replication strategy taking a replication factor per datacenter and placing replicas on distinct racks where possible, so a rack or datacenter outage does not remove every copy.

QUORUM consistency

Requires floor(RF/2)+1 replica acknowledgements, which is two when RF is three. Combining QUORUM reads and writes makes read plus write exceed RF, giving strong consistency.

Hinted handoff

The coordinator stores writes destined for an unreachable replica and replays them when it returns. max_hint_window_in_ms defaults to three hours, after which repair is required.

LOCAL_QUORUM

Quorum computed only over replicas in the coordinator datacenter. It avoids cross-datacenter latency in multi-datacenter clusters while still tolerating local replica failure.

SizeTieredCompactionStrategy

Default compaction strategy merging similarly sized SSTables. It is write-optimized but needs roughly fifty percent free disk and keeps obsolete or deleted data until compaction occurs.

LeveledCompactionStrategy

Keeps SSTables in levels of fixed size (sstable_size_in_mb, default 160) with non-overlapping ranges, so reads touch few files. Suits update-heavy and read-heavy tables.

TimeWindowCompactionStrategy

Groups SSTables into time windows and compacts only within each window, so TTL-expired time-series data is discarded as whole SSTables without rewriting older data.

Key cache

Maps partition keys to SSTable offsets, skipping index lookups on repeated reads. Its size defaults to the smaller of five percent of heap and 100 MB.

nodetool repair -full

Compares all replica data using Merkle trees and streams differences. Cassandra 4.0 restored full repair as the default after incremental repair proved difficult to reconcile reliably.

gc_grace_seconds

Window, defaulting to 864000 seconds (ten days), during which tombstones are retained so every replica learns about deletes. Repair must complete inside this window.

nodetool snapshot

Creates hard links to current SSTables in the snapshots directory, an almost instantaneous point-in-time copy. auto_snapshot also takes one before truncate or drop.

Commit log archiving

archive_command copies commit log segments to external storage, allowing point-in-time restore when replayed together with an SSTable snapshot taken earlier.

PasswordAuthenticator

cassandra.yaml authenticator that stores role credentials as hashes in the system_auth keyspace. That keyspace needs a replication factor above one in every datacenter.

CassandraAuthorizer

Authorizer class recording GRANT permissions per role in system_auth. It must be paired with an authenticator, because it enforces permissions but never verifies identity itself.

Internode encryption

Configured in server_encryption_options. internode_encryption set to none, dc, rack, or all controls which node-to-node links use TLS with the configured keystore and truststore.

Role-based access control

Cassandra 3.0 and later treat users as roles that hold permissions and can be granted to other roles. Superuser roles bypass permission checks entirely.

nodetool drain

Flushes memtables to disk then stops accepting writes and stops gossip. Run it before stopping a node for maintenance or an upgrade to avoid data loss.

Schema disagreement

Nodes reporting different schema versions. Inspect with nodetool describecluster, then restart the divergent nodes or run nodetool resetlocalschema to force agreement.

upgradesstables

After a major version upgrade, nodetool upgradesstables rewrites existing SSTables into the new format so the node uses current read and compaction paths.

Dropped messages

nodetool tpstats counters for requests rejected because a thread pool queue filled. Sustained drops indicate overload and typically precede read or write timeouts.