MongoDB Production Engineering

Design documents around how data is read, not just how it's conceptually related — MongoDB rewards read-optimized, denormalized shapes.

$jsonSchema validation rules enforce required fields and types at the database level, without giving up per-document flexibility entirely.

An index on { customerId: 1, createdAt: -1 } serves queries filtering on customerId alone, or customerId + createdAt range — but not createdAt alone. Order matters (the ESR rule: Equality, Sort, Range).

{ partialFilterExpression: { status: 'pending' } } — index only the documents a query actually needs, keeping the index small.

If every field a query needs is present in the index itself, MongoDB can answer without fetching the document at all.

Placing $match (and $sort, when it can use an index) as early as possible lets MongoDB use indexes and shrinks the working set for every later stage.

A $lookup join scans the foreign collection per input document unless it can use an index on the foreign field — always index the field a $lookup joins on.

A big gap between documents examined and documents returned is the #1 sign a query needs a better index.

COLLSCAN means no usable index was found; IXSCAN means an index narrowed the search before fetching documents.

primary, primaryPreferred, secondary, secondaryPreferred, or nearest — controls which member(s) can serve a read, trading recency for read scalability.

The oplog is a capped collection — if it's too small, a secondary that falls behind (e.g. after maintenance) may not be able to catch up and needs a full resync.

Needs high cardinality and even write distribution. A monotonically increasing key (like a timestamp) creates a 'hot shard' that absorbs all new writes.

The balancer moves chunks between shards to keep data evenly distributed — this consumes I/O and should be scheduled during low-traffic windows for large migrations.

Waits for a majority of voting replica set members to acknowledge before confirming a write — survives a primary failure immediately after the write. It is the implicit default in MongoDB 5.0+ (with documented exceptions, such as some arbiter-containing topologies, and configurable cluster-wide defaults); earlier versions default to w: 1.

Guarantees a read only returns data acknowledged by a majority — avoiding reads of data that could later be rolled back after a failover.

Logs slow operations above a configurable threshold — the first place to look for a specific slow query in production.

Live view of connections, memory, and in-flight operations — the equivalent of Postgres' pg_stat_activity.

rs.printSecondaryReplicationInfo() shows how far behind each secondary is — critical for read-preference correctness and failover readiness.

An embedded array that grows without limit (e.g. comments on a post) eventually approaches the 16 MB document size limit and slows every read/write of that document.

Scatter-gather queries that hit every shard, when a targeted query using the shard key would have hit just one.

A secondary taken down for maintenance falls off the back of the oplog and requires a full, expensive resync instead of catching up.