MongoDB Subset and Computed Patterns

Q: MongoDB Subset and Computed Patterns

The **subset pattern** keeps a hot slice of a growing array (last 10 reviews, recent 20 comments) in the primary document and moves the tail to a separate collection — so the main document stays small and fast to load. The **computed pattern** stores pre-calculated values (averages, totals, rankings) alongside the raw data so reads never recompute on the fly. Both patterns trade **a bit of write complexity for dramatically cheaper reads**, and both exist because MongoDB's **16MB document cap** plus working-set economics punish unbounded embedding. Use subset when 90% of reads want only the recent slice; use computed when the same aggregation runs thousands of times per read.

Intermediate·2 min read·

MongodbEbayShopifyHboAirbnb

Fast recall

Quick revision

Core answer

The subset pattern keeps a hot slice of a growing array (last 10 reviews, recent 20 comments) in the primary document and moves the tail to a separate collection — so the main document stays small and fast to load. The computed pattern stores pre-calculated values (averages, totals, rankings) alongside the raw data so reads never recompute on the fly. Both patterns trade a bit of write complexity for dramatically cheaper reads, and both exist because MongoDB's 16MB document cap plus working-set economics punish unbounded embedding. Use subset when 90% of reads want only the recent slice; use computed when the same aggregation runs thousands of times per read.

Interview-ready guide

Interview answer

Frame both patterns

Subset keeps only the hot slice in the main document, tail in a side collection
Computed stores pre-calculated aggregates so reads never recompute
Both trade a bit of write complexity for much cheaper reads

Show the idioms

$push with $slice and $sort for subset
$inc for counters; background $merge for heavier aggregates
Index the hot slice fields if you query them

Call out the trade-offs

Write-heavy workloads can flip the math
Staleness is acceptable for most aggregates; design refresh cadence
Compound failures — wrap in transaction or run nightly reconcile

Wrap up

Both exist because the 16MB limit plus working-set economics punish unbounded embedding
Combine them — products with recentReviews slice plus ratingAvg computed field is a canonical shape

03 · Deep dive