Free audit

View Audit Scope
JusDBAmazon DynamoDB Consulting · Single-Table PK/SK · 1,000 WCU Sharding · AWS FinOps

Master Single-Table DynamoDB, eliminate 1,000 WCU hot keys, and cut RCU/WCU waste.

Amazon DynamoDB delivers predictable single-digit millisecond latency at any scale, provided your schema respects physical partition mechanics. When teams lift relational tables straight into DynamoDB, `Scan` filters, dense `ALL`-projected GSIs, and unsharded partition keys hit the hard `1,000 WCU` per-partition ceiling and inflate AWS bills. JusDB’s principal NoSQL architects design composite `PK`/`SK` Single-Table schemas, remediate GSI back-pressure, right-size On-Demand vs Provisioned + `Standard-IA` capacity, and execute live S3/Streams table migrations.
Our engineering team is in India (IST, UTC+5:30). Or request a free audit of one database instance.
  • 15 minutes P1 response target under a 24x7 support agreement
  • 1,000 WCU hot-key sharding & GSI back-pressure fixes
  • 35-65% typical AWS bill reduction via GSI pruning & RC
  • 24 supported database technologies

DynamoDB Architecture Matrix

Composite PK/SK Overloading, Item Collections & Sparse GSIs

ARCHITECTURE · PK/SK

DynamoDB does not execute relational joins. Modeling one table per entity forces application code to run N+1 network round-trips, whereas a well-engineered Single-Table Design retrieves parent-child hierarchies in a single `Query` API call.

We map every application access pattern before writing code, overload generic `PK`/`SK` attributes (`TENANT#id`, `ORDER#ts`), and design sparse Global Secondary Indexes with `KEYS_ONLY` or `INCLUDE` projections.

Scope a Single-Table design review

Engineered across Amazon DynamoDB, AWS, PostgreSQL, MongoDB & Cassandra estates

  • Innoviti
  • Blitz
  • Freecharge
  • Nango
  • Recochain
  • Screen Facts
  • Nexora
  • WeCredit

01 · Production Failure & Cost Modes

Six DynamoDB bottlenecks we resolve inside physical partitions, GSIs, and Streams.

Switching a table to On-Demand capacity does not raise the 1,000 WCU / 3,000 RCU physical partition ceiling or stop GSI write back-pressure. Here is how we diagnose and fix the underlying data model.

Exception / Telemetry SignalArchitectural Root CauseHow We VerifyProduction Remediation
ProvisionedThroughputExceededException · Hot partition keyA single partition key value exceeds the hard physical partition limit of `1,000 WCU` or `3,000 RCU`, triggering throttling even when the table is configured for On-Demand capacity or has ample aggregate provisioned throughput.Enable CloudWatch Contributor Insights on the base table and inspect `ThrottledRequests`, `WriteThrottleEvents`, and `ReadThrottleEvents` per key.Introduce synthetic write-sharding suffixes (`TENANT#42#SHARD_0`..`SHARD_15`) calculated via hash modulo or random distribution, and fan out reads with `BatchGetItem` or parallel `Query` calls.
GSI_WRITE_BACKPRESSURE · Base table writes throttled by indexA Global Secondary Index partitioned on a low-cardinality attribute (such as `status = 'PENDING'` or `date = '2026-11-09'`) saturates a single GSI physical partition (`1,000 WCU`) and throttles writes on the base table.Compare base table `WriteThrottleEvents` against GSI `OnlineIndexThrottleEvents` and `WriteThrottleEvents` in CloudWatch.Convert dense status GSIs into sparse indexes (only populate the GSI PK attribute while an item is active, deleting the attribute on completion) and shard GSI partition keys.
SCAN_FILTER_TRAP · Unbounded RCU burn on filtered queriesUsing `Scan` or `Query` with a `FilterExpression` evaluates the filter only after reading and billing RCUs for every item in the underlying key range (up to 1 MB per page).Compare `ScannedCount` against `Count` in SDK response telemetry (`ReturnConsumedCapacity: INDEXES`).Push selective filter predicates into composite sort keys (`SK` `begins_with` or `BETWEEN` ranges) or dedicated sparse GSIs so `ScannedCount` equals `Count`.
GSI_PROJECTION_AMPLIFICATION · 4x to 6x inflated WCU billDeclaring 4 or 5 Global Secondary Indexes with `ProjectionType: ALL` on large items (> 4 KB) forces every single attribute update on the base table to write full item copies across every GSI.Audit `aws dynamodb describe-table` index projections, average item byte sizes, and per-index `ConsumedWriteCapacityUnits`.Restrict GSIs to `KEYS_ONLY` or minimal `INCLUDE` attribute lists, split mutable counters into separate items within the same item collection, and consolidate redundant GSIs via overloaded `GSI1PK`/`GSI1SK` keys.
STREAMS_POISON_PILL · 24-hour shard retention expiration riskAn unhandled schema exception inside an AWS Lambda consumer attached to DynamoDB Streams retries the same batch endlessly until the 24-hour stream retention window expires and events are lost.Monitor Lambda `IteratorAge` in CloudWatch alongside stream consumer error counts.Enable `BisectBatchOnFunctionError`, configure `MaximumRetryAttempts` and an SQS/SNS `OnFailure` Dead-Letter Queue (DLQ), and enforce idempotent consumer processing.
DAX_CACHE_MISMATCH · Paying for idle or write-thrashing DAXDeploying DynamoDB Accelerator (DAX) in front of write-heavy or `TransactGetItems`/`ConsistentRead` workloads bypasses or thrashes the item/query cache while billing continuous EC2-equivalent node hours.Check DAX `ItemCacheHits`, `ItemCacheMisses`, `QueryCacheHits`, and CPU utilization against application read consistency flags.Retire DAX where workloads are write-heavy or strongly consistent, or replace with targeted ElastiCache (Valkey/Redis) look-aside caching for aggregated objects.
Evaluation Vector
JusDB DBRE
In-House DBALegacy AgencyDeveloper Generalist
Single-Table Design & Composite Key (PK/SK) ModelingNormalizes complex relational entity hierarchies into optimized single-table models using composite partition and sort keys (PK/SK), sparse indexes, and GSI overloading to serve the OLTP access patterns at single-digit millisecond p99 latencies.Recreates traditional relational schemas across dozens of discrete tables, executing expensive application-side JOINs that multiply latency, compute overhead, and consumed request units across requests.Treats DynamoDB like a relational SQL database; deploys one table per entity and defaults to unindexed, unbounded table Scan operations that stall during peak user traffic.Defines ad-hoc key schemas without composite hierarchies or access pattern planning, frequently hitting partition limits and forcing full table scans for basic entity lookups.
Synthetic Suffix Sharding (1,000 WCU Partition Cap Elimination)Identifies high-velocity write keys and designs deterministic or randomized synthetic suffix sharding (e.g. PK#001–PK#050) with scatter-gather parallel queries, permanently bypassing the strict 1,000 WCU per physical partition ceiling.Attempts to fix 1,000 WCU partition throttles by globally increasing provisioned table WCU, multiplying AWS cloud bills while hot-partition throttling persists unchanged.Blames DynamoDB engine constraints and bolts on slow external caching layers that introduce stale data inconsistencies and cache invalidation race conditions.Unaware of the 1,000 WCU physical partition ceiling; directs high-velocity time-series or event streams to single monolithic partition keys, causing immediate P1 write throttling.
On-Demand vs Provisioned Autoscaling & Reserved Capacity SizingAnalyzes CloudWatch consumed capacity distributions, burst profiles, and traffic histograms to calibrate auto-scaled provisioned capacity and commit to 1- or 3-year Reserved Capacity where the usage pattern supports it.Leaves mission-critical tables in default on-demand mode permanently for operational simplicity, paying up to a 7x pricing premium on steady-state predictable workloads.Configures rigid static provisioned capacity thresholds with disabled autoscaling, causing widespread throttling during sudden bursts and heavy over-spend during off-peak hours.Switches capacity modes reactively based on sporadic billing alarms without calculating consumed capacity curves or commitment breakeven thresholds.
Global Secondary Index (GSI) Over-Provisioning & Write AmplificationAudits query projection attributes to engineer KEYS_ONLY or INCLUDE GSIs, preventing table-wide write throttling from unprojected GSI backpressure and minimizing storage write amplification.Configures ALL attribute projections on every secondary index by default, multiplying table write amplification, storage volume, and cross-region replication expenses by 3x–5x.Provisions redundant, overlapping GSIs for minor query variations without monitoring consumed index write capacity or index backpressure metrics.Omits GSIs entirely or provisions inverted indexes that suffer from extreme partition key skew, throttling primary table write transactions during bursts.
Global Tables Multi-Region Active-Active Replication & OCCArchitects active-active multi-region Global Tables with deterministic conflict resolution, Optimistic Concurrency Control (OCC) attribute versioning, and real-time cross-region replication latency telemetry.Enables Global Tables without OCC versioning or idempotency controls, resulting in silent data overwrites, race conditions, and last-writer-wins data corruption during concurrent cross-region updates.Builds fragile custom application-tier replication microservices, leading to desynchronized regional datasets, split-brain states, and cascading reconciliation failures.Deploys Global Tables across distant geographies without evaluating cross-region data transfer pricing amplification or eventual consistency read replication delays.
DynamoDB Streams & Poison-Pill Resilient Event-Driven Consumer PipelinesImplements DynamoDB Streams with Kinesis Adapter and AWS Lambda consumers featuring exponential backoff, Dead Letter Queues (DLQ), and bisect-batch-on-error, so failed records are retried or captured in the DLQ instead of being dropped.Attaches unmonitored Lambda functions directly to DynamoDB Streams without DLQs or error bisecting; a single malformed poison-pill record halts entire shard processing for hours.Avoids native CDC Streams and runs scheduled cron scripts polling tables via scans, incurring massive RCU consumption and introducing multi-minute pipeline data lag.Writes unhandled stream consumer functions that crash on schema variations, causing shard iterator expiration and permanent, silent event stream data loss.

02 · Consulting Workstreams

End-to-end engineering across Single-Table design, FinOps, and Global Tables.

Every engagement delivers concrete access-pattern matrices, Terraform/CDK definitions, and benchmarked SDK query layers.

01 · SINGLE-TABLE MODELING

Access-Pattern Mapping, Single-Table PK/SK & Item Collections

We catalog every read and write access pattern alongside latency and concurrency targets before designing your table. We model composite `PK` and `SK` hierarchies, overload generic `GSI1PK`/`GSI1SK` through `GSI3PK`/`GSI3SK` indexes, and use `TransactWriteItems` for uniqueness constraints.

PK/SK overloading · Sparse GSIs · Item Collections · TransactWriteItems

02 · HOT-KEY & SHARDING

1,000 WCU / 3,000 RCU Hot-Partition Remediation & Write Sharding

We eliminate `ProvisionedThroughputExceededException` and GSI write back-pressure using CloudWatch Contributor Insights forensics, synthetic write-sharding suffixes (`PK#00`..`PK#15`), scatter-gather read patterns, and sparse index lifecycle hygiene.

Contributor Insights · Synthetic key suffixing · GSI back-pressure fix

03 · FINOPS & TABLE CLASSES

On-Demand vs Provisioned FinOps, Standard-IA & GSI Rightsizing

We audit RCU/WCU consumption curves across your AWS accounts. We switch predictable workloads from On-Demand to Provisioned Auto-Scaling with Reserved Capacity, prune `ALL` GSI projections to `KEYS_ONLY`/`INCLUDE`, transition cold tables to `Standard-IA`, and automate TTL S3 archival.

Provisioned + Reserved Capacity · Standard-IA · KEYS_ONLY · TTL archival

04 · STREAMS & CDC

DynamoDB Streams, Kinesis Data Streams & Zero-ETL Analytics

We architect resilient event pipelines over DynamoDB Streams and Kinesis Data Streams with shard-bisecting Lambda error handling, DLQs, and EventBridge Pipes. For search and analytics, we configure Zero-ETL integrations with Amazon OpenSearch Service, Redshift, or ClickHouse.

DynamoDB Streams · BisectBatchOnFunctionError · EventBridge Pipes · Zero-ETL

05 · GLOBAL TABLES & DR

Multi-Region Global Tables, Cross-Region PITR & S3 Export/Import

We design active-active Global Tables across AWS regions with idempotent last-writer-wins conflict safeguards, or cost-optimized single-region architectures backed by Point-in-Time Recovery (PITR, 1 to 35 days) and automated `ExportTableToPointInTime` S3 snapshots.

Global Tables · 35-day PITR · ExportTableToPointInTime · ImportTable

06 · LIVE TABLE MIGRATIONS

Relational-to-DynamoDB & Table-to-Table Live Re-Keying Pipelines

Because a DynamoDB primary key cannot be altered in place (and AWS DMS only supports DynamoDB as a target, not a source), we execute live table-to-table re-keying using S3 PITR Export + `ImportTable` paired with DynamoDB Streams delta catch-up and shadow dual-reads.

S3 PITR Export + ImportTable · Streams delta sync · Dual-read cutover

03 · Responsibility Boundary & Response Targets

Where AWS managed storage ends and JusDB DynamoDB ownership begins.

DynamoDB has no servers to patch, which means almost every latency spike, throttling outage, or cost overrun originates in partition key design, GSI projection choices, or stream consumer configuration.

Shared Responsibility Matrix (Amazon DynamoDB)

Physical Storage Nodes, 3-AZ Paxos Replication & AWS Control Plane

Amazon Web Services (AWS)

AWS manages underlying SSD storage partitions, synchronous 3-AZ replication within a region, and hardware replacement.

JusDB DynamoDB Architects

We architect VPC Endpoints, KMS Customer Managed Keys, resource-based IAM policies, and 35-day PITR + cross-account S3 backup vaults.

Single-Table PK/SK Schema, GSI Projections & 1,000 WCU Partition Ceiling

Amazon Web Services (AWS)

Enforces the `1,000 WCU` / `3,000 RCU` per-partition ceiling, the 400 KB item limit, and throttles hot keys or back-pressured GSIs.

JusDB DynamoDB Architects

We design composite `PK`/`SK` access patterns, write-sharded suffixes, and sparse `KEYS_ONLY`/`INCLUDE` GSIs so partitions never hot-spot.

Capacity Mode (On-Demand vs Provisioned), Auto-Scaling & Table Class FinOps

Amazon Web Services (AWS)

Bills per million On-Demand requests or per provisioned RCU/WCU hour, plus storage and full-item GSI replication writes.

JusDB DynamoDB Architects

We right-size Auto-Scaling bounds, model 1-year/3-year Reserved Capacity, convert cold tables to `Standard-IA`, and eliminate `Scan` queries.

DynamoDB Streams Consumers, Global Tables & 24x7 Incident Response (Where Contracted)

Amazon Web Services (AWS)

Retains DynamoDB Streams records for 24 hours and replicates Global Tables asynchronously using last-writer-wins timestamps.

JusDB DynamoDB Architects

We harden stream consumers against poison-pill stalls, monitor `ReplicationLatency`, and provide 24x7 DBA response under a 24x7 support agreement.

Target Incident Response Matrix

P1 to P4 Engineering Targets (15m / 30m / 60m / 90m)

When a hot key or GSI back-pressure wave throttles checkout traffic, the page goes to a senior AWS database engineer. Initial-response targets are 2 to 4 hours during covered business hours, Mon–Fri 09:00–18:00 IST (UTC+5:30). Where your service agreement includes 24x7 on-call, the initial-response target for P1 incidents is 15 minutes. Targets are not resolution guarantees.

P1 · CriticalTarget: 15 minutes

Production database down, primary unreachable, data corruption risk, or active failover event.

24x7 on-call (under your service agreement) · Direct PagerDuty, Opsgenie, or phone page plus incident bridge

P2 · HighTarget: 30 minutes

Severe latency spike, replication lag climbing rapidly, connection exhaustion, or degraded HA node.

24x7 on-call (under your service agreement) · On-call alert plus shared Slack or Teams channel

P3 · MediumTarget: 60 minutes

Non-critical slow query regression, staging issue, backup warning, or replication check alert.

24x7 on-call (under your service agreement) · Shared Slack or Teams channel or support ticket

P4 · LowTarget: 90 minutes

Configuration question, index advisory, planned schema review, or general operational request.

24x7 on-call (under your service agreement) · Shared Slack or Teams channel or support ticket

Experiencing active `ProvisionedThroughputExceededException` throttling or stream lag? Open an emergency escalation →

04 · Interactive Partition & Single-Table Simulator

See how write-sharded PKs and sparse GSIs eliminate throttling.

Compare a throttled table suffering from a hot partition key and dense GSI back-pressure against a write-sharded Single-Table design and a multi-region Global Tables architecture.

Amazon DynamoDB Partition & Global Tables Simulator

Application and routing

App SDK to Hot Partition Key (TENANT#enterprise) to Dense GSI (status = PENDING) Back-Pressure

THROTTLED · 1,000 WCU CAP BREACHED
  • Base Table Partition #1 (PK = TENANT#enterprise)

    1,450 Write Ops/sec Against 1,000 WCU Ceiling

    ProvisionedThroughputExceededException

  • GSI-Status (GSI_PK = 'PENDING', Projection = ALL)

    All New Orders Funnel into One GSI Partition

    GSI BACK-PRESSURE HALTING BASE WRITES

  • Order Listing Endpoint (Scan + FilterExpression)

    Reading 12,000 Items to Return 25 Rows

    480 RCU PER CALL · 95ms p99 LATENCY

Even on an On-Demand table, routing more than 1,000 WCU to a single partition key or a dense `status = PENDING` GSI key triggers immediate throttling and halts base-table ingestion.

05 · Representative Deliverable

Contributor Insights forensics and write-sharded Single-Table items.

Every DynamoDB consulting engagement produces concrete AWS CLI diagnostics, access-pattern item templates, and Terraform/CDK definitions engineered to keep every physical partition under 1,000 WCU.

  • 01Deterministic Write-Sharding: Appends `#SHARD_00`..`#SHARD_07` to high-velocity tenant keys so burst writes spread across 8 physical partitions.
  • 02Self-Cleaning Sparse GSI: Populates `GSI1PK` and `GSI1SK` only while an order is `IN_FLIGHT` and removes both attributes upon completion to keep the operational index tiny.
  • 03Automated TTL Hygiene: Uses native epoch `expiresAtEpoch` attributes to age out transient items at zero WCU cost.

dynamodb-hot-key-and-single-table-pattern.json

# 1. AWS CLI: Inspect table & GSI throttling + replication latency (last 6 hours)
aws cloudwatch get-metric-statistics \
  --namespace AWS/DynamoDB \
  --metric-name WriteThrottleEvents \
  --dimensions Name=TableName,Value=prod-commerce-single-table \
  --start-time 2026-11-09T00:00:00Z \
  --end-time   2026-11-09T06:00:00Z \
  --period 300 \
  --statistics Sum

# 2. Enable CloudWatch Contributor Insights to isolate hot PK/SK keys
aws dynamodb update-contributor-insights \
  --table-name prod-commerce-single-table \
  --contributor-insights-action ENABLE

# 3. Single-Table item pattern: Write-sharded PK + Sparse GSI1 (only present while active)
{
  "PK": { "S": "TENANT#t_9481#SHARD_03" },
  "SK": { "S": "ORDER#2026-11-09T11:20:04Z#ord_7712" },
  "entityType": { "S": "ORDER" },
  "orderStatus": { "S": "IN_FLIGHT" },
  "amountCents": { "N": "149900" },
  "GSI1PK": { "S": "IN_FLIGHT_BUCKET#03" },
  "GSI1SK": { "S": "2026-11-09T11:20:04Z#ord_7712" },
  "expiresAtEpoch": { "N": "1825728000" }
}
# Note: When orderStatus transitions to COMPLETED, REMOVE GSI1PK, GSI1SK
# so completed orders immediately drop out of the sparse operational index.

06 · Engagement Models & Scope

Clear commercial paths from schema & FinOps audits to live table migrations.

Engage JusDB for a fixed-scope Single-Table and FinOps audit, a turnkey live table re-keying sprint, or ongoing 24x7 AWS database reliability operations under a 24x7 support agreement.

Engineering SprintRecommended

Single-Table Design, Live Table Re-Keying & Relational Migration

End-to-end design and execution of a new Single-Table schema, live table-to-table re-partitioning via S3 PITR Export + `ImportTable` + DynamoDB Streams delta catch-up, or migration from RDS/Aurora/MongoDB/Cassandra into DynamoDB.

Included in scope

  • Access-pattern specification, item collection schema design, and TypeScript/Go/Java/Python SDK data layer
  • Live re-keying of production DynamoDB tables using S3 Export/Import + Streams CDC + dual-read verification
  • Global Tables multi-region active-active rollout or Streams/Kinesis + OpenSearch Zero-ETL integration
Fixed-Scope Audit

DynamoDB Single-Table, Hot-Key & RCU/WCU FinOps Audit

A comprehensive engineering review of your DynamoDB tables: access-pattern vs `PK`/`SK` alignment, Contributor Insights hot-partition forensics, GSI projection bloat, and On-Demand vs Provisioned + Reserved Capacity + `Standard-IA` cost modeling.

Included in scope

  • Complete audit of base table and GSI key distributions against the 1,000 WCU / 3,000 RCU ceiling
  • GSI consolidation and projection rightsizing (`ALL` to `KEYS_ONLY`/`INCLUDE` and sparse index conversion)
  • Table-by-table FinOps spreadsheet: On-Demand vs Provisioned Auto-Scaling, Reserved Capacity & Standard-IA
Ongoing Retainer

Managed DynamoDB SRE, Cost Governance & On-Call DBA Support

Ongoing architectural governance, throttling prevention, stream consumer reliability, and 24x7 incident response under a 24x7 support agreement for high-scale AWS estates.

Included in scope

  • P1 critical acknowledgment target: 15 minutes under a 24x7 support agreement
  • 24x7 monitoring of `WriteThrottleEvents`, `OnlineIndexThrottleEvents`, and Lambda `IteratorAge` under a 24x7 support agreement
  • Quarterly Reserved Capacity & Standard-IA FinOps tuning plus cross-account PITR restore drills

Procurement and security

Ready for legal and security review.

All engagements include NDA execution, least-privilege IAM AssumeRole access, and reproducible Terraform/CDK artifacts.
  • Mutual NDA before any access
  • Named individual accounts only
  • Read-only access by default
  • ISO 27001 & SOC 2 Type II: aligned, certification in progress
  • Measured AWS FinOps Savings

    Typical 35% to 65% monthly bill reduction via GSI pruning, Provisioned RC & Standard-IA

  • Support Coverage

    2 to 4 hours (During covered business hours) · 15 minutes under a 24x7 support agreement

  • Security & Access Governance

    ISO 27001 & SOC 2 Type II: aligned, certification in progress · Least-privilege IAM AssumeRole inside your AWS Organization

  • Commercial Models

    Fixed-fee schema & FinOps audits, milestone migration sprints, or monthly SRE retainers

07 · Technical & FinOps FAQ

Common questions on Amazon DynamoDB consulting.

Direct answers from our AWS database architects on Single-Table Design, the 1,000 WCU partition limit, GSI write back-pressure, Standard-IA FinOps, and live S3/Streams table migrations.

Still have a question?

Ask it on a scoping call. No production access needed.

Book a 30-min scoping call (opens Google Calendar in a new tab)

When is Single-Table Design worth the complexity in Amazon DynamoDB, and when should we avoid it?

Single-Table Design (packing multiple related entities such as `Tenant`, `Order`, and `OrderItem` into one DynamoDB table using generic `PK` and `SK` attributes) is ideal for latency-critical OLTP microservices with well-defined access patterns that otherwise require multi-table round-trips. Because all items sharing a `PK` sit together in the same sorted item collection, a single `Query` (`PK = 'ORDER#123'`) fetches the order header and all line items in under 5 milliseconds. However, Single-Table Design makes ad-hoc BI queries and per-entity access control harder; for microservices with independent lifecycles or heavy ad-hoc analytics, we often pair domain-scoped multi-table designs with Zero-ETL streaming into OpenSearch, Redshift, or ClickHouse.

Why does our On-Demand DynamoDB table still throw ProvisionedThroughputExceededException?

On-Demand mode eliminates table-level capacity planning, but it does not change the physics of an individual storage partition. Every single DynamoDB physical partition has a hard ceiling of `1,000 Write Capacity Units (WCU)` and `3,000 Read Capacity Units (RCU)` per second. If a flash sale, celebrity account, or status bucket routes more than 1,000 1-KB writes per second to the exact same partition key value, DynamoDB cannot split a single partition key across multiple partitions and will return `ProvisionedThroughputExceededException`. We solve this by adding deterministic or random write-sharding suffixes (`PK#00`..`PK#15`) and using scatter-gather reads.

How can a Global Secondary Index (GSI) throttle writes on our base DynamoDB table?

DynamoDB updates Global Secondary Indexes asynchronously, pulling from an internal buffer between the base table and the GSI. If a GSI uses a low-cardinality partition key (such as `status = 'NEW'` or a coarse date bucket) and exceeds `1,000 WCU` on a single GSI partition, or if a provisioned GSI lacks sufficient WCU capacity, that internal buffer fills up. Once the buffer saturates, DynamoDB applies immediate back-pressure and rejects `PutItem`/`UpdateItem` calls on the base table itself. We prevent GSI back-pressure by designing sparse GSIs (deleting the GSI attribute once an item leaves the active workflow) and sharding high-velocity GSI keys.

How do we choose between On-Demand and Provisioned capacity (plus Reserved Capacity) to lower our AWS bill?

On a unit-per-request basis, On-Demand requests cost significantly more than fully utilized Provisioned throughput, especially when paired with 1-year or 3-year DynamoDB Reserved Capacity (which discounts provisioned RCU/WCU by over 50% to 75%). We keep spiky, unpredictable, or low-volume tables on On-Demand (using configurable On-Demand maximum throughput limits to prevent runaway bills), and switch steady or diurnal high-volume tables to Provisioned capacity with Target Tracking Auto-Scaling and Reserved Capacity.

When does switching a table to DynamoDB Standard-IA actually save money?

The `Standard-Infrequent Access (Standard-IA)` table class lowers storage prices by roughly 60% compared to the `Standard` table class, but increases read and write request pricing by roughly 25%. As a rule of thumb, switching to `Standard-IA` lowers your total monthly bill when storage accounts for more than 40% to 50% of that table's total DynamoDB cost (including its GSIs). For truly cold archival data that rarely needs millisecond retrieval, exporting via TTL or S3 PITR Export to Parquet in Amazon S3 is even more cost-effective.

How do you change a DynamoDB partition key or migrate a live table without downtime?

Because DynamoDB does not allow modifying a table's `PK` or `SK` schema after creation, and because AWS DMS only supports DynamoDB as a migration target (never as a source), live re-keying requires a dual-pipeline strategy. First, we enable Point-in-Time Recovery (PITR) and DynamoDB Streams on the source table. Second, we record a timestamp `T0` and run `ExportTableToPointInTime` to Amazon S3, transforming and loading the historical snapshot into the new target table via `ImportTable` or Spark/Step Functions. Third, we replay all DynamoDB Stream events since `T0` using idempotent conditional writes (`ConditionExpression`) until lag reaches sub-second parity, verify via dual-reads, and flip application writes.

When should we deploy DynamoDB Global Tables vs a single-region table with cross-region backups?

DynamoDB Global Tables replicate writes across multiple AWS regions automatically for active-active multi-region availability and local read latency, but every write is billed in every participating region (`rWCU`), and concurrent updates to the same item in two regions resolve via last-writer-wins (LWW) wall-clock timestamps. If your compliance or RTO target requires sub-minute multi-region failover, we engineer Global Tables with region-pinned write routing or CRDT-safe attributes. If your RTO allows 30 to 60 minutes, a single-region table with cross-region S3 PITR exports avoids multi-region write amplification.

What response times do you provide for production Amazon DynamoDB incidents?

Under a 24x7 support agreement, our target initial response times are: P1 (Critical) within 15 minutes; P2 (High) within 30 minutes; P3 (Medium) within 60 minutes; P4 (Low) within 90 minutes. During standard business-hours consulting retainers, our response target is 2 to 4 hours. Every escalation goes straight to a senior AWS & NoSQL database reliability engineer.

Talk to a Principal AWS & DynamoDB Architect

Ready to eliminate hot partitions and cut your DynamoDB RCU/WCU bill?

Share your table access patterns, CloudWatch throttling metrics, or monthly DynamoDB cost breakdown for a pragmatic engineering assessment.