Cut Your MongoDB Costs by 79% with Shape‑First Optimizations

Protect your database from future fires to avoid high-scale capital loss at your Series A stage Protect your database from future fires to avoid high-scale capital loss at your Series A stage Disclaimer: The following is a fictional case study used to communicate best practices for MongoDB schema design, performance tuning, and cost optimization Disclaimer: The following is a fictional case study used to communicate best practices for MongoDB schema design, performance tuning, and cost optimization Disclaimer: The following is a fictional case study used to communicate best practices for MongoDB schema design, performance tuning, and cost optimization The Day the Bill Went Nuclear The call came through at 2:17. 2:17 Atlas caused another unsolicited production cluster scale-up that resulted in an M60 machine at a monthly cost of $15k. The board wanted to know why burn increased by 20% while M60 serves as an expensive $15 k/month system. M60 $15k $15 k/month I opened the profiler: I opened the profiler: db.system.profile.aggregate([ { $match: { millis: { $gt: 100 } } }, { $group: { _id: { op: "$op", ns: "$ns", query: "$command.filter" }, n: { $sum: 1 }, avgMs: { $avg: "$millis" } }}, { $sort: { n: -1 } }, { $limit: 10 } ]).pretty(); db.system.profile.aggregate([ { $match: { millis: { $gt: 100 } } }, { $group: { _id: { op: "$op", ns: "$ns", query: "$command.filter" }, n: { $sum: 1 }, avgMs: { $avg: "$millis" } }}, { $sort: { n: -1 } }, { $limit: 10 } ]).pretty(); Each dashboard widget required the offender to pull in a total of 1.7 GB every minute before rendering. The sheer quantity of memory usage created mountain peaks in the graph that resembled Everest. 1.7 GB The M30 servers now operate with one of these clusters. The solution did not entail an increase in shards. Three common flaws known as shape crimes existed in the codebase before elimination. shape crimes Crime Scene Investigation 2.1 N + 1 Query Tsunami This is recognized as an anti-pattern—when ordering one set of orders requires running N separate queries to retrieve order lines. // Incorrect: Orders + 1 000 extra queries const orders = await db.orders.find({ userId }).toArray(); for (const o of orders) { o.lines = await db.orderLines.find({ orderId: o._id }).toArray(); } // Incorrect: Orders + 1 000 extra queries const orders = await db.orders.find({ userId }).toArray(); for (const o of orders) { o.lines = await db.orderLines.find({ orderId: o._id }).toArray(); } Hidden taxes Hidden taxes Meter Why it spikes Compute 1 000 cursors = 1 000 context switches Storage I/O 1 000 index walks + 1 000 doc deserialisations Network Each round‑trip eats ~1 ms RTT + TLS handshake overhead Meter Why it spikes Compute 1 000 cursors = 1 000 context switches Storage I/O 1 000 index walks + 1 000 doc deserialisations Network Each round‑trip eats ~1 ms RTT + TLS handshake overhead Meter Why it spikes Meter Meter Why it spikes Why it spikes Compute 1 000 cursors = 1 000 context switches Compute Compute Compute 1 000 cursors = 1 000 context switches 1 000 cursors = 1 000 context switches Storage I/O 1 000 index walks + 1 000 doc deserialisations Storage I/O Storage I/O Storage I/O 1 000 index walks + 1 000 doc deserialisations 1 000 index walks + 1 000 doc deserialisations Network Each round‑trip eats ~1 ms RTT + TLS handshake overhead Network Network Network Each round‑trip eats ~1 ms RTT + TLS handshake overhead Each round‑trip eats ~1 ms RTT + TLS handshake overhead Refactor (4 lines): Refactor (4 lines): // Success: Single round‑trip, 1 read unit per order db.orders.aggregate([ { $match: { userId } }, { $lookup: { from: "orderLines", localField: "_id", foreignField: "orderId", as: "lines" }}, { $project: { lines: 1, total: 1, ts: 1 } } ]); // Success: Single round‑trip, 1 read unit per order db.orders.aggregate([ { $match: { userId } }, { $lookup: { from: "orderLines", localField: "_id", foreignField: "orderId", as: "lines" }}, { $project: { lines: 1, total: 1, ts: 1 } } ]); Latency p95 dropped from 2 300 ms to 160 ms. Latency p95 dropped from 2 300 ms to 160 ms. 2 300 ms 160 ms Atlas read‑ops: 101 → 1. That’s 99 % off—no coupon code needed. 101 → 1. 2.2 Unbounded Query Firehose “But we have to show the full click history!” “But we have to show the full click history!” Sure—just not in a single cursor. // Failure: Streams 30 months of data through the API gateway db.events.find({ userId }).toArray(); // Failure: Streams 30 months of data through the API gateway db.events.find({ userId }).toArray(); Fix: hard‑cap the batch and project only the fields you render. Fix db.events.find( { userId, ts: { $gte: ISODate(new Date() - 1000*60*60*24*30) } }, { _id: 0, ts: 1, page: 1, ref: 1 } // projection ).sort({ ts: -1 }).limit(1_000); db.events.find( { userId, ts: { $gte: ISODate(new Date() - 1000*60*60*24*30) } }, { _id: 0, ts: 1, page: 1, ref: 1 } // projection ).sort({ ts: -1 }).limit(1_000); Then let Mongo clean up behind you: // 90‑day sliding window db.events.createIndex({ ts: 1 }, { expireAfterSeconds: 60*60*24*90 }); // 90‑day sliding window db.events.createIndex({ ts: 1 }, { expireAfterSeconds: 60*60*24*90 }); One fintech client cut their storage bill 72 % overnight simply by adding TTLs. One fintech client cut their storage bill 72 % overnight simply by adding TTLs. 2.3 Jumbo‑Document Money Pit Mongo caps documents at 16 MB, but anything over 256 KB is already a red flag. { "_id": "...", "type": "invoice", "customer": { /* 700 kB */ }, "pdf": BinData(0,"..."), // 4 MB binary "history": [ /* 1 200 delta rows */ ], "ts": ISODate() } { "_id": "...", "type": "invoice", "customer": { /* 700 kB */ }, "pdf": BinData(0,"..."), // 4 MB binary "history": [ /* 1 200 delta rows */ ], "ts": ISODate() } Why it hurts Why it hurts The whole doc is paged in even if you read one field. WiredTiger can’t store as many docs per page → lower cache hit ratio. Index entries are huge → bloom filter misses → more disk seeks. The whole doc is paged in even if you read one field. The whole doc is paged in even if you read one field. WiredTiger can’t store as many docs per page → lower cache hit ratio. WiredTiger can’t store as many docs per page → lower cache hit ratio. Index entries are huge → bloom filter misses → more disk seeks. Index entries are huge → bloom filter misses → more disk seeks. Solution: schema‑by‑access‑pattern: Solution schema‑by‑access‑pattern graph TD Invoice[(invoices &lt;2 kB)] -->|ref| Hist[history &lt;1 kB * N] Invoice -->|ref| Bin[pdf‑store (S3/GridFS)] graph TD Invoice[(invoices &lt;2 kB)] -->|ref| Hist[history &lt;1 kB * N] Invoice -->|ref| Bin[pdf‑store (S3/GridFS)] Small invoice metas stay hot; BLOBS in S3 cost $0.023/GB‑month instead of NAND‑grade Atlas SSDs. Small invoice metas stay hot; BLOBS in S3 cost $0.023/GB‑month instead of NAND‑grade Atlas SSDs. Four More Shape Crimes You’re Probably Guilty Of Low‑cardinality index head ({ type: 1, ts: -1 })—re‑order it to { userId: 1, ts: -1 }. $regex starts‑with on non‑indexed field—string scan from hell. findOneAndUpdate queue—document‑level locking bottleneck; use Redis/Kafka. skip + large offset pagination—Mongo must count every skipped doc; switch to range (ts, _id) cursors. Low‑cardinality index head ({ type: 1, ts: -1 })—re‑order it to { userId: 1, ts: -1 }. Low‑cardinality index head { type: 1, ts: -1 } { userId: 1, ts: -1 } $regex starts‑with on non‑indexed field—string scan from hell. $regex starts‑with findOneAndUpdate queue—document‑level locking bottleneck; use Redis/Kafka. findOneAndUpdate queue skip + large offset pagination—Mongo must count every skipped doc; switch to range (ts, _id) cursors. skip + large offset range (ts, _id) 4 · Cost Anatomy 101 “But Atlas says reads are cheap!” “But Atlas says reads are cheap!” Let’s do the math. Metric Value Unit cost Monthly cost Reads (3 k/s) 7.8 B $0.09 / M $702 Writes (150 /s) 380 M $0.225 / M $86 Data transfer 1.5 TB $0.25 / GB $375 Storage (2 TB) $0.24 / GB $480 Metric Value Unit cost Monthly cost Reads (3 k/s) 7.8 B $0.09 / M $702 Writes (150 /s) 380 M $0.225 / M $86 Data transfer 1.5 TB $0.25 / GB $375 Storage (2 TB) $0.24 / GB $480 Metric Value Unit cost Monthly cost Metric Metric Value Value Unit cost Unit cost Monthly cost Monthly cost Reads (3 k/s) 7.8 B $0.09 / M $702 Reads (3 k/s) Reads (3 k/s) 7.8 B 7.8 B $0.09 / M $0.09 / M $702 $702 $702 Writes (150 /s) 380 M $0.225 / M $86 Writes (150 /s) Writes (150 /s) 380 M 380 M $0.225 / M $0.225 / M $86 $86 $86 Data transfer 1.5 TB $0.25 / GB $375 Data transfer Data transfer 1.5 TB 1.5 TB $0.25 / GB $0.25 / GB $375 $375 $375 Storage (2 TB) $0.24 / GB $480 Storage (2 TB) Storage (2 TB) $0.24 / GB $0.24 / GB $480 $480 $480 Total: $1,643. $1,643. Apply the fixes: Reads fall 70 % → $210 Transfer falls 80 % → $75 Storage falls 60 % → $192 Reads fall 70 % → $210 Transfer falls 80 % → $75 Storage falls 60 % → $192 New bill: $564. That’s one mid‑level engineer or runway till Q4—you choose. New bill: $564. That’s one mid‑level engineer or runway till Q4—you choose. New bill: $564. or 48‑Hour Rescue Sprint (Battle‑tested Timeline) Hour Action Tool Wins 0‑2 Turn on profiler (slowms = 50). Mongo shell Surface top 10 slow ops. 2‑6 Rewrite N + 1 into $lookup. VS Code + Jest tests 90 % fewer reads. 6‑10 Add projections & limit to unbounded finds. API layer RAM steady; API 4× faster. 10‑16 Break jumbo docs → metas + GridFS/S3. Scripted ETL Working set fits in RAM. 16‑22 Drop/replace low‑cardinality indexes. Compass Disk shrinks; cache hits ↑. 22‑30 Create TTLs, month‑partition cold data, enable Online Archive. Atlas UI 60 % storage saved. 30‑36 Add Grafana panels: cache hit %, scan:ix ratio, eviction rate. Prometheus Visual early warnings. 36‑48 Load‑test with k6 k6 + Atlas metrics Confirm p95 < 150 ms @ 2× load. Hour Action Tool Wins 0‑2 Turn on profiler (slowms = 50). Mongo shell Surface top 10 slow ops. 2‑6 Rewrite N + 1 into $lookup. VS Code + Jest tests 90 % fewer reads. 6‑10 Add projections & limit to unbounded finds. API layer RAM steady; API 4× faster. 10‑16 Break jumbo docs → metas + GridFS/S3. Scripted ETL Working set fits in RAM. 16‑22 Drop/replace low‑cardinality indexes. Compass Disk shrinks; cache hits ↑. 22‑30 Create TTLs, month‑partition cold data, enable Online Archive. Atlas UI 60 % storage saved. 30‑36 Add Grafana panels: cache hit %, scan:ix ratio, eviction rate. Prometheus Visual early warnings. 36‑48 Load‑test with k6 k6 + Atlas metrics Confirm p95 < 150 ms @ 2× load. Hour Action Tool Wins Hour Hour Action Action Tool Tool Wins Wins 0‑2 Turn on profiler (slowms = 50). Mongo shell Surface top 10 slow ops. 0‑2 0‑2 Turn on profiler (slowms = 50). Turn on profiler (slowms = 50). slowms = 50 Mongo shell Mongo shell Surface top 10 slow ops. Surface top 10 slow ops. 2‑6 Rewrite N + 1 into $lookup. VS Code + Jest tests 90 % fewer reads. 2‑6 2‑6 Rewrite N + 1 into $lookup. Rewrite N + 1 into $lookup. $lookup VS Code + Jest tests VS Code + Jest tests 90 % fewer reads. 90 % fewer reads. 6‑10 Add projections & limit to unbounded finds. API layer RAM steady; API 4× faster. 6‑10 6‑10 Add projections & limit to unbounded finds. Add projections & limit to unbounded finds. limit API layer API layer RAM steady; API 4× faster. RAM steady; API 4× faster. 10‑16 Break jumbo docs → metas + GridFS/S3. Scripted ETL Working set fits in RAM. 10‑16 10‑16 Break jumbo docs → metas + GridFS/S3. Break jumbo docs → metas + GridFS/S3. Scripted ETL Scripted ETL Working set fits in RAM. Working set fits in RAM. 16‑22 Drop/replace low‑cardinality indexes. Compass Disk shrinks; cache hits ↑. 16‑22 16‑22 Drop/replace low‑cardinality indexes. Drop/replace low‑cardinality indexes. Compass Compass Disk shrinks; cache hits ↑. Disk shrinks; cache hits ↑. 22‑30 Create TTLs, month‑partition cold data, enable Online Archive. Atlas UI 60 % storage saved. 22‑30 22‑30 Create TTLs, month‑partition cold data, enable Online Archive. Create TTLs, month‑partition cold data, enable Online Archive. Atlas UI Atlas UI 60 % storage saved. 60 % storage saved. 30‑36 Add Grafana panels: cache hit %, scan:ix ratio, eviction rate. Prometheus Visual early warnings. 30‑36 30‑36 Add Grafana panels: cache hit %, scan:ix ratio, eviction rate. Add Grafana panels: cache hit %, scan:ix ratio, eviction rate. Prometheus Prometheus Visual early warnings. Visual early warnings. 36‑48 Load‑test with k6 k6 + Atlas metrics Confirm p95 10? → Refactor. Cursor returns > 1 000 docs? → Paginate. TTL on every event/store table? (Yes/No) Any index where cardinality 1 % total ops? → Optimize or cache. Largest doc ÷ median > 10? → Refactor. Largest doc ÷ median > 10? → Refactor. Cursor returns > 1 000 docs? → Paginate. Cursor returns > 1 000 docs? → Paginate. TTL on every event/store table? (Yes/No) TTL on every event/store table? (Yes/No) Any index where cardinality 1 % total ops? → Optimize or cache. Profiler slowops > 1 % total ops? → Optimize or cache. If primary cache hits remain under 90% it is wise to separate collections or add additional RAM memory post fixes. If primary cache hits remain under 90% it is wise to separate collections or add additional RAM memory post fixes. Place the checklist on your laptop with adhesive glue after laminating it for printing. Place the checklist on your laptop with adhesive glue after laminating it for printing. Why Shape Beats Indexes MongoDB’s query planner does a cost‑based search across candidate plans. The cost vector includes: workUnits = ixScans + fetches + sorts + #docs returned workUnits = ixScans + fetches + sorts + #docs returned Indexes only reduce ixScans. Bad shape inflates fetches and sorts, which often dominate. Example: ixScans fetches sorts db.logs.find( { ts: { $gte: start, $lt: end }, level: "error" } ).sort({ level: 1, ts: -1 }); db.logs.find( { ts: { $gte: start, $lt: end }, level: "error" } ).sort({ level: 1, ts: -1 }); Index { level: 1, ts: -1 } does not help planner avoid fetching every document when it adds a predicate to an unmentioned array field in your projections. Net result: 20 k fetches for 200 hits. Index should precede shape operations in daily operations. { level: 1, ts: -1 } Live Metrics You Should Be Watching (Grafana PromQL) # WiredTiger cache hit ratio (rate(wiredtiger_blockmanager_blocks_read[1m]) / (rate(wiredtiger_blockmanager_blocks_read[1m]) + rate(wiredtiger_blockmanager_blocks_read_from_cache[1m])) ) 10 % misses for 5 m. # Docs scanned vs returned rate(mongodb_ssm_metrics_documents[1m]{state="scanned"}) / rate(mongodb_ssm_metrics_documents[1m]{state="returned"}) > 100 # Docs scanned vs returned rate(mongodb_ssm_metrics_documents[1m]{state="scanned"}) / rate(mongodb_ssm_metrics_documents[1m]{state="returned"}) > 100 If you scan 100× more docs than you return, you’re burning money. If you scan 100× more docs than you return, you’re burning money. Hands‑On: Thin‑Slice Migration Script Need to crack a 1‑TB events collection into clicks, views, logins without downtime? Use the double‑write / backfill pattern. events clicks views logins double‑write / backfill // 1. Add trigger const changeStream = db.events.watch([], { fullDocument: 'updateLookup' }); changeStream.on('change', ev => { const dest = db[`${ev.fullDocument.type}s`]; dest.insertOne({ ...ev.fullDocument }); }); // 2. Backfill historical in chunks let lastId = ObjectId("000000..."); while (true) { const batch = db.events.find({_id: {$gt: lastId}}).sort({_id: 1}).limit(10_000); if (!batch.hasNext()) break; const docs = batch.toArray(); docs.forEach(d => db[`${d.type}s`].insertOne(d)); lastId = docs[docs.length - 1]._id; } // 1. Add trigger const changeStream = db.events.watch([], { fullDocument: 'updateLookup' }); changeStream.on('change', ev => { const dest = db[`${ev.fullDocument.type}s`]; dest.insertOne({ ...ev.fullDocument }); }); // 2. Backfill historical in chunks let lastId = ObjectId("000000..."); while (true) { const batch = db.events.find({_id: {$gt: lastId}}).sort({_id: 1}).limit(10_000); if (!batch.hasNext()) break; const docs = batch.toArray(); docs.forEach(d => db[`${d.type}s`].insertOne(d)); lastId = docs[docs.length - 1]._id; } Zero downtime, minimal extra storage (thanks to TTL), everyone sleeps. Zero downtime, minimal extra storage (thanks to TTL), everyone sleeps. When Sharding Is the Answer Is According to the rule of thumb you should shard only if you verify one of these conditions comes true after optimizing your database: one The system operates with a working set representing more than 80 percent of RAM regardless of the cache performance rate. The system generates more than 15 thousand operations per second in its peak write performance when using one primary server. Your main priorities should be maintaining sub-70-millisecond multi-region latency because high AWS billing costs do not represent your critical concern. The system operates with a working set representing more than 80 percent of RAM regardless of the cache performance rate. The system operates with a working set representing more than 80 percent of RAM regardless of the cache performance rate. The system generates more than 15 thousand operations per second in its peak write performance when using one primary server. The system generates more than 15 thousand operations per second in its peak write performance when using one primary server. Your main priorities should be maintaining sub-70-millisecond multi-region latency because high AWS billing costs do not represent your critical concern. Your main priorities should be maintaining sub-70-millisecond multi-region latency because high AWS billing costs do not represent your critical concern. The decision should be simple when the conditions do not match these rules. The decision should be simple when the conditions do not match these rules. Case Study Wrap‑Up Metric Before After Δ RAM footprint 120 GB 36 GB −70 % Reads/sec 6 700 900 −86 % Storage (hot) 2.1 TB 600 GB −71 % p95 latency 1.9 s 140 ms −92 % Atlas cost / mo. $15 284 $3 210 −79 % Metric Before After Δ RAM footprint 120 GB 36 GB −70 % Reads/sec 6 700 900 −86 % Storage (hot) 2.1 TB 600 GB −71 % p95 latency 1.9 s 140 ms −92 % Atlas cost / mo. $15 284 $3 210 −79 % Metric Before After Δ Metric Metric Before Before After After Δ Δ RAM footprint 120 GB 36 GB −70 % RAM footprint RAM footprint 120 GB 120 GB 36 GB 36 GB −70 % −70 % Reads/sec 6 700 900 −86 % Reads/sec Reads/sec 6 700 6 700 900 900 −86 % −86 % Storage (hot) 2.1 TB 600 GB −71 % Storage (hot) Storage (hot) 2.1 TB 2.1 TB 600 GB 600 GB −71 % −71 % p95 latency 1.9 s 140 ms −92 % p95 latency p95 latency 1.9 s 1.9 s 140 ms 140 ms −92 % −92 % Atlas cost / mo. $15 284 $3 210 −79 % Atlas cost / mo. Atlas cost / mo. $15 284 $15 284 $3 210 $3 210 −79 % −79 % No shards, no major code freeze, just ruthless shape surgery. No shards, no major code freeze, just ruthless shape surgery. Takeaway: Debt vs. Death Spiral The requirement to deliver quickly stands as mandatory, but maintaining poor quality work belongs to voluntary debt accumulation. Cloud provider charges you compound interest that accumulates with a yearly percentage rate of 1000 % on your unpaid debt. We examined the high‑interest credit cards of MongoDB as they represent the five shape crimes we studied. Shedding these debts within your current sprint period will result in grateful technical and financial statements. You need to open the profiler to work with the $lookup pistons while adding TTL dust, and then deploy your project in lean mode. Your board and your dev team and your pager at 02:17 will obtain quality rest. $lookup Proceed with the refactoring of your code until the next autoscaling incident happens.