MongoDB DBA Associate cheat sheet

Replica set minimum for production high availability is three voting members; an arbiter provides a vote but no data and is unsuitable when the primary plus one data-bearing secondary cannot survive a failure.

A sharded cluster requires exactly these components: config server replica set (CSRS), one or more shard replica sets, mongos routers, and drivers connecting only to mongos.

Config servers must run as a replica set (CSRS) since MongoDB 3.4; legacy three-mirror (SCCC) config servers are unsupported in modern versions.

mongos is stateless and holds no persistent data; deploy at least two mongos instances and place them near application servers to avoid a single router as a failure point.

Bind IP defaults to 127.0.0.1; a node reachable by peers must set net.bindIp to the specific interface address or a comma-separated list, not 0.0.0.0 in hardened environments.

For a replica set, all members must share the same major/minor version family during rolling upgrades; the FCV (featureCompatibilityVersion) gates which features are active.

Election-related topology rule: an even number of voting members adds no fault tolerance over the next lower odd number; use arbiters or non-voting members to adjust votes without adding data-bearing nodes.

Read preference 'primaryPreferred' fails over reads to a secondary only when no primary is available; 'nearest' selects by measured latency among eligible members, not by geographic proximity alone.

A majority is floor(votingMembers/2)+1; a primary must be able to see a majority to remain primary, otherwise it steps down.

Default election trigger: a member starts an election after electionTimeoutMillis (default 10000 ms) without hearing from the primary; heartbeats run every 2 seconds.

Priority 0 members can never become primary; hidden members must have priority 0 and are invisible to client read preference; delayed members apply an oplog delay via secondaryDelaySecs.

Votes are capped at 7; members beyond the seventh voting member must be configured with votes:0, and changing votes requires reconfig with the correct force option when a majority is unreachable.

w:majority write concern requires acknowledgment from a majority of data-bearing voting members; with w:1 a write can be rolled back if the primary fails before replication.

A rollback occurs when a node rejoins and has oplog entries the new primary never had; MongoDB writes the removed documents to rollback files for manual reconciliation.

rs.stepDown(seconds) forces the primary to relinquish; rs.freeze(seconds) prevents a member from seeking election; both are used to control failover timing during maintenance.

catchUpTakeoverDelayMillis (default 30000 ms) makes a higher-priority secondary wait for a better-positioned candidate before calling an election after a primary stepdown.

Chunks default to 128 MB; the balancer migrates chunks, not documents, and migration rounds are triggered by chunk-count imbalance across shards.

Hashed shard keys give even write distribution but scatter reads for range queries; ranged keys support targeted range reads but risk hotspotting on monotonically increasing values.

A shard key is immutable on documents before MongoDB 5.0; from 5.0 onward documents can be moved between shards with updateOne using a full shard key in the filter.

Targeted queries include the full shard key; scatter-gather queries omit it and fan out to all shards, so shard key choice dominates query cost.

The balancer can be disabled with sh.stopBalancer() or per-window scheduling via sh.setBalancerState and the balancer window in the config database settings collection.

Zones associate shard key ranges with named shards; each zone must have at least one shard and each shard belongs to at most one zone; chunks need a matching zone range to migrate.

mongos caches routing metadata in the config database; stale routing after chunk migration is retried transparently, but long-running transactions can pin older metadata.

A unique index on a sharded collection must contain the shard key as a prefix; otherwise the index cannot be enforced cluster-wide.

WiredTiger uses document-level concurrency control; readers see a consistent snapshot without blocking writers, and writes are buffered and checkpointed every 60 seconds by default.

The journal (WiredTiger log) provides crash durability; writeConcern j:true forces journal acknowledgment, and j:false allows loss of writes since the last journal flush.

WiredTiger cache defaults to 50% of (RAM minus 1 GB) or 256 MB, whichever is larger; the cacheSizeGB setting caps this and eviction begins around 80% dirty or 95% total cache use.

The oplog is a capped collection in the local database; its size is set by replication.oplogSizeMB, and a too-small oplog causes secondaries to fall off and require full resync.

Embedding favors one-to-few and read-together data; referencing favors one-to-many with unbounded growth or independently accessed entities, avoiding the 16 MB document limit.

The 16 MB BSON document limit constrains array growth; the bucket pattern and outlier pattern split large logical entities across documents.

Schema validation uses $jsonSchema with validationLevel (strict/moderate) and validationAction (error/warn); collMod changes validation on an existing collection.

Index intersection is limited; compound indexes ordered by equality, sort, then range fields usually outperform relying on the planner to intersect single-field indexes.

SCRAM-SHA-256 is the default authentication mechanism; internal member authentication uses keyfile or x.509, and enabling authorization without internal auth breaks replica set communication.

Built-in roles include read, readWrite, dbAdmin, clusterAdmin, userAdmin, root, and backup/restore; custom roles combine privileges with resource patterns such as {db:'app', collection:'*'}.

The localhost exception allows creating the first user only when no users exist and the node is accessed from localhost; it disappears after the first user is created.

TLS is enabled with net.tls.mode requireTLS plus certificateKeyFile; clients may present certificates for x.509 auth, and net.tls.allowConnectionsWithoutCertificates controls optional client certs.

Encryption at rest uses the WiredTiger encrypted storage engine with a key management provider: local keyfile for testing, KMIP or cloud KMS for production; the master key never resides in the data files.

Client-side field level encryption (CSFLE) encrypts specified fields before they reach the server; queryable encryption additionally permits equality or range queries on encrypted fields.

Auditing is an enterprise feature configured under auditLog with destination file, syslog, or console; the filter document selects which events are recorded.

Authentication restrictions: authenticationRestrictions on a user limits source IP ranges or CIDR blocks, and db.auth() in mongosh switches the authenticated user for the session.

mongodump captures BSON plus index definitions but not the oplog unless --oplog is used; restoring with --oplogReplay yields a point-in-time consistent snapshot only when the dump was taken with --oplog.

Filesystem snapshots require journaling enabled and a consistent state: either stop writes or use db.fsyncLock() before the snapshot and db.fsyncUnlock() after.

mongorestore --drop removes existing collections before restoring; without --drop, restored documents are merged into existing collections and can create duplicates.

A logical dump of a sharded cluster must go through mongos; per-shard dumps are inconsistent because chunk migrations may be in flight.

Point-in-time recovery combines a base backup with oplog replay to a timestamp; the oplog must extend back to the base backup time or recovery is impossible.

Restore testing is a control, not a formality: a backup is only valid once a restore into a separate environment has been verified against expected document counts and indexes.

For large datasets, physical backup (filesystem or Ops Manager/Cloud Manager snapshot) is faster than mongodump and preserves the exact on-disk format.

Recovery time objective drives topology: a three-member replica set tolerates one failure, and cross-region members protect against site loss at the cost of write latency under w:majority.

serverStatus exposes opcounters, connections, wiredTiger cache metrics, and globalLock queues; currentOp and $currentOp show in-flight operations, and killOp terminates a specific opid.

The database profiler levels are 0 (off), 1 (slow ops above slowms), and 2 (all ops); profiling writes to the system.profile capped collection and adds overhead.

explain('executionStats') reports totalKeysExamined, totalDocsExamined, and nReturned; a healthy index scan has totalDocsExamined close to nReturned.

A CO COLLSCAN stage in explain indicates a collection scan; IXSCAN with a high keys-to-docs ratio suggests a poorly selective index or a filter not covered by the index.

Replication lag is measured by comparing optimeDate of primary and secondary via rs.status(); a growing lag with a small oplog risks the secondary falling off and needing resync.

The slow query log threshold is set by operationProfiling.slowms or the slowms parameter; logLevel above 0 increases verbosity and should be temporary.

db.stats() and db.collection.stats() report storageSize, indexSize, and wiredTiger block-manager metrics used to size disks and detect index bloat.

mongotop reports per-collection read/write time and mongostat reports per-second opcounters, faults, and queue depths; both are point-in-time and best sampled over intervals.

Rolling upgrade order: upgrade secondaries one at a time, step down the primary so a secondary takes over, upgrade the former primary, then raise featureCompatibilityVersion last.

Never downgrade the FCV below the version required by stored data; setFCV to the previous version must precede any binary downgrade and only works within supported downgrade paths.

After a major upgrade, run the compatibility check: FCV stays at the old value until explicitly raised, so new features remain unavailable until setFeatureCompatibilityVersion runs.

Index builds on replica sets build on all members; on sharded clusters, unique index builds require a stop-the-world procedure or are coordinated by mongos depending on version.

Compaction reclaims disk space from WiredTiger fragmentation; it is run per-node with db.runCommand({compact:...}) and blocks that database's operations during execution.

Adding a member uses rs.add() or rs.addArb(); the new member performs an initial sync, and a large dataset may need maintenance windows or a seed from a filesystem snapshot.

Disk-full handling: WiredTiger stops accepting writes when the volume is full; monitor disk free space and keep the journal on the same volume sized for oplog plus data growth.

A runbook for a failed primary documents: detect via rs.status(), confirm majority, verify new primary optime, re-add the repaired node, and validate replication lag returns to baseline.

mongodump --gzip --archive=backup.gz writes a single compressed archive; mongorestore --gzip --archive=backup.gz restores it, and --nsFrom/--nsTo rename namespaces during restore.

mongodump --oplog captures oplog entries during the dump for a consistent point-in-time restore with mongorestore --oplogReplay; it is unsupported for sharded clusters.

mongosh supports scripting with load(), EJSON for extended types, and --eval for one-liners; db.getSiblingDB() switches databases within a session.

mongosh connection strings accept SRV records (mongodb+srv://), authMechanism, and tls options; the shell respects readPreference for secondary reads.

Ops Manager and Cloud Manager provide automated backup with point-in-time restore, monitoring alerts, and rolling upgrade orchestration for self-managed deployments.

The Ops Manager agent (mongodb-mms-automation-agent) runs on each host and performs deployments and backups; the backup daemon stores snapshots in a blockstore or S3 bucket.

mongoexport and mongoimport handle JSON/CSV and are for data exchange, not backup; they lose BSON type fidelity and do not capture indexes.

mongostat and mongotop ship with the database tools and accept --rowcount and --discover to poll a URI for all members of a replica set or sharded cluster.