TanStack Query Best Practices

Expert patterns for TanStack Query in modern React applications with Next.js App Router and Server Components.

NEVER Do With TanStack Query

• NEVER use a singleton QueryClient on the server - Creates data leakage between users and race conditions. Each request must get its own isolated QueryClient instance to prevent cached data from one user appearing for another.
• NEVER synchronize query data to useState - Background refetches, invalidations, and optimistic updates all modify the cache. Local state copies become stale immediately, causing "my save didn't work" bugs. Use query data directly or derive with useMemo.
• NEVER put queries inside list item components - Creates N observers for N items, causing O(n) iteration on every cache update. 200 list items calling useQuery creates 200 network requests and 200 observers. Hoist queries to parent components.
• NEVER use unstable query keys - Arrays with non-guaranteed order, temporal queries with Date.now(), or object keys without deterministic serialization create infinite cache entries. Keys must be stable and deterministic.
• NEVER skip enabled guards for dependent queries - Firing queries with undefined parameters creates garbage cache entries like ['tracks', undefined] and wastes network requests before real data arrives.
• NEVER ignore AbortController signals - Without query cancellation support, unmounted components leave in-flight requests running, wasting bandwidth and potentially updating stale cache entries.
• NEVER use optimistic updates for high-stakes or external mutations - Life-critical operations, audit trail systems, and mutations triggered by external events need pessimistic updates to ensure UI matches server state.
• NEVER assume structural sharing is free - For datasets >1000 items updating frequently, structural sharing's O(n) deep equality checks become CPU overhead. Disable with structuralSharing: false for large, frequently-changing data.
• NEVER skip onSettled in optimistic updates - onSettled is your cleanup guarantee even if onError throws. Without it, UI can be left in corrupted state when error handler fails. Always pair onMutate with onSettled for resource cleanup and cache consistency.
• NEVER assume cache invalidation is synchronous - invalidateQueries triggers background refetches which can race with optimistic updates. Use cancelQueries in onMutate to prevent background refetches from overwriting your optimistic changes before the mutation completes.
• NEVER use setQueryData without structural comparison - Directly setting cache data bypasses structural sharing and breaks referential equality optimizations. Wrap in updater function to preserve references for unchanged portions: setQueryData(key, (old) => ({ ...old, changed: value })) instead of setQueryData(key, newValue).
• NEVER forget to handle hydration mismatches - Server-rendered data may differ from client expectations (timestamps, user-specific data, randomized content). Use suppressHydrationWarning on containers or ensure deterministic server/client rendering with stable timestamps and consistent data sources.

Before Using TanStack Query, Ask

State Classification

• Is this server state or client state? TanStack Query manages server state (API data, database records, external system state). UI state (modals, themes, form drafts) belongs in Zustand or useState.
• Does this data change after initial render? Static reference data might not need TanStack Query's refetching machinery. Consider if simpler alternatives suffice.

Cache Strategy

• How fresh does this data need to be? Lookup tables can have 1-hour staleTime. Real-time tracking needs 5-second staleTime with refetchInterval. Match configuration to business requirements.
• What's the query lifecycle? Frequently-accessed data needs higher gcTime. One-time detail views can have aggressive garbage collection.

Observer Economics

• How many components will subscribe to this query? >10 observers on a single cache entry suggests hoisting queries to parent. >100 observers indicates architectural issues.
• Am I creating N queries or 1 query with N observers? List items should receive props from parent query, not call individual useQuery hooks.

How to Use

This skill uses progressive disclosure to minimize context usage. Load references based on your scenario:

Scenario 1: Setting Up Query Client

MANDATORY - READ ENTIRE FILE: Read query-client-setup.md (~125 lines) and server-integration.md (~151 lines) completely for server/client setup patterns. Do NOT Load other references for initial setup.

Copy assets/query-client.ts for production-ready configuration.

Scenario 2: Building Query Hooks

1. MANDATORY: Read query-keys.md (~151 lines) for key factory setup
2. If using server components: Read server-integration.md
3. Do NOT Load mutations-and-updates.md unless implementing mutations

Use decision tables below for configuration values.

Scenario 3: Implementing Mutations

MANDATORY - READ ENTIRE FILE: Read mutations-and-updates.md (~345 lines) completely. Reference patterns-and-pitfalls.md for rollback patterns. Do NOT Load caching-strategy.md for basic CRUD mutations.

Scenario 4: Debugging Performance Issues

1. First, check Observer Count Thresholds table below (lines 121-129)
2. If observer count >50: Read patterns-and-pitfalls.md
3. If large dataset issues: Read fundamentals.md for structural sharing
4. Do NOT Load all references - diagnose first, then load targeted content

Scenario 5: Multi-Layer Caching Strategy

MANDATORY: Read caching-strategy.md (~198 lines) for unified Next.js use cache + TanStack Query + HTTP cache patterns. Do NOT Load if only using client-side TanStack Query.

Query Configuration Decision Matrix

Mutation Pattern Selection

Query Key Architecture

Use hierarchical factories for consistent invalidation:

// Recommended structure
export const keys = {
  all: () => ['domain'] as const,
  lists: () => [...keys.all(), 'list'] as const,
  list: (filters: string) => [...keys.lists(), filters] as const,
  details: () => [...keys.all(), 'detail'] as const,
  detail: (id: string) => [...keys.details(), id] as const,
};

// Invalidation examples
queryClient.invalidateQueries({ queryKey: keys.all() }); // Invalidate everything
queryClient.invalidateQueries({ queryKey: keys.lists() }); // Invalidate all lists
queryClient.invalidateQueries({ queryKey: keys.detail(id) }); // Invalidate one item

Key stability rules:

• Deterministic serialization (sort arrays before joining)
• No temporal values (Date.now(), random IDs)
• Type consistency (don't mix '1' and 1)
• Stable object shapes (use sorted keys or serialize)

Server-Client Integration Pattern

Unified invalidation strategy:

1. Use same key factories for both server and client caches
2. Server mutations call updateTag(keys.detail(id).tag)
3. Client mutations call queryClient.invalidateQueries({ queryKey: keys.detail(id) })
4. Both caches stay synchronized with same hierarchy

Observer Count Thresholds

Diagnosis:

1. Open TanStack Query DevTools in development
2. Find cache entries with high observer counts
3. Search codebase for useQuery calls with those keys
4. Refactor to parent components or shared cache entries

Query Hook Patterns

Common Error Patterns and Fixes

Troubleshooting Decision Tree

Performance Issues

Step 1: Check observer count in DevTools (use thresholds at lines 136-145)

• >100 observers → Immediate refactor: hoist queries to parent component, distribute data via props
• 51-100 observers → Refactor: hoist queries or use select to subscribe to data subsets
• <50 observers → Issue is elsewhere, continue to Step 2

Step 2: Check data size and update frequency

• >1000 items + frequent updates → Disable structural sharing: structuralSharing: false (see fundamentals.md for details)
• Large payloads (>500KB) → Check network tab, consider pagination or infinite queries
• Fast updates (<1s interval) → Lower staleTime or use refetchInterval, verify cache strategy

Step 3: Check React DevTools Profiler

• Look for unnecessary re-renders in components using query data
• Verify select function isn't recreated on every render (use useCallback)
• Check if derived data should use useMemo instead of inline transformation
• Profile component render times to identify bottlenecks

Network Issues

Flaky connections:

• Configure retry logic: retry: 3, retryDelay: (attemptIndex) => Math.min(1000 * 2 ** attemptIndex, 30000)
• See query-client-setup.md for production retry configuration

Token refresh needed:

• Implement auth interceptor pattern in queryFn wrapper
• Use queryClient.setQueryDefaults() for global auth headers
• See patterns-and-pitfalls.md for token refresh patterns

Race conditions:

• Review invalidation timing: use cancelQueries before setQueryData
• Check if optimistic updates compete with background refetches
• Verify mutation onMutate uses await cancelQueries({ queryKey })

Hydration Issues

SSR mismatch (hydration error in console):

• Add suppressHydrationWarning to container element
• Normalize data: ensure server and client produce identical output (stable timestamps, sorted arrays)
• Check if user-specific data is being rendered server-side

Client-server data drift:

• Verify revalidateTag timing on server mutations
• Check if server cache (Next.js use cache) is stale while client cache is fresh
• Use initialData from server props: initialData: serverData, refetchOnMount: false

HydrationBoundary not working:

• Verify client component boundaries: HydrationBoundary must wrap 'use client' components
• Check if dehydratedState is being serialized correctly from server
• Ensure shouldDehydrateQuery includes queries you want to hydrate

Freedom Calibration

Calibrate guidance specificity to mutation risk:

The test: "If the agent makes a mistake, what's the consequence?"

• Server singleton mistake → Data leakage between users (critical security issue)
• Observer count mistake → Performance degradation (medium impact)
• staleTime tuning → Suboptimal freshness (low impact)

Important Notes

• Query keys are hashed deterministically - ['tracks', '1'] and ['tracks', 1] create different cache entries
• Query keys must be JSON-serializable for cache persistence across page reloads and hydration
• shouldDehydrateQuery with pending status enables streaming without await in server components
• HydrationBoundary must wrap client components only - server components bypass the boundary
• revalidateTag vs updateTag matters: revalidateTag uses stale-while-revalidate, updateTag invalidates immediately
• Background refetches run even when no components are mounted if gcTime hasn't expired
• Structural sharing runs twice when using select: once on raw data, once on transformed data
• select only runs on successfully cached data — it is never called in error states; put validation and error throwing in queryFn
• cancelQueries in onMutate is critical - background refetches can overwrite optimistic updates
• Context returned from onMutate is passed to onError and onSettled for rollback state