Collections Best Practices

Q: When is OrderedDict still justified?

move_to_end LRU patterns. Regular dict handles insertion order otherwise.

Q: Should I always use slots?

__slots__ for millions of small fixed-field objects. Premature for typical app records.

Q: Counter vs defaultdict(int)?

Counter adds multiset operators and most_common - prefer for frequency work.

Q: How avoid defaultdict serialization surprises?

Convert dict(defaultdict) before JSON - factory not serialized.

Practical rules for choosing and using Python's built-in containers without performance traps or readability debt.

How to Use This List

Apply during code review on data-heavy modules.
Revisit when profiling shows unexpected O(n) behavior.
Pair with type hints (dict[str, int]) for clearer contracts.
Use as onboarding checklist for data pipeline code.

A - Selection

Pick dict for keyed lookup, set for membership, list for ordered sequences. Document non-obvious choices.
Use deque for FIFO/LIFO queues instead of list.pop(0). O(1) vs O(n) at front.
Use heapq for priority scheduling; bisect for sorted insertions. Not sorted() on every insert.
Reach for pandas 2.2+ or Polars 1.x at tabular scale, not nested dicts of lists. Column ops and joins.
Prefer Counter and defaultdict over manual key-exists boilerplate. Fewer branches, clearer intent.

B - Correctness

Never use mutable lists or dicts as dict keys or set elements. Use tuple, frozenset, or frozen dataclass.
Shallow-copy nested structures with copy.deepcopy when independence required. dict.copy() shares inner lists.
Treat set iteration order as unspecified. Use sorted(s) for stable display only.
Validate uniqueness before inverting dict {v: k}. Colliding values overwrite silently.
Use get or try/except KeyError intentionally. d[k] when key must exist; get when optional.

C - Performance and Memory

Build set once for repeated membership in loops. Avoid O(n) in list in hot paths.
Use generator expressions feeding sum/any/max for large streams. Skip materializing million-item lists.
Consider array.array or bytes for compact numeric/binary buffers. Not list of ints for wire protocols.
Profile before swapping list for exotic structure. Small n favors simplicity.
Avoid repeated keys() snapshot unless mutating during iteration. list(d) when deleting while looping.

D - Readability

Use dict/set comprehensions for transforms with optional filters. Clearer than manual loops.
Name types in hints: dict[str, list[Event]], not bare dict. Self-documenting structures.
Use NamedTuple or dataclass instead of positional tuple indexing for records. row[3] obscures meaning.
Prefer | merge for immutable config overlays (3.9+). base | overrides reads left-to-right.
Order-preserving dedupe: list(dict.fromkeys(items)). Idiomatic one-liner.

E - Serialization and APIs

JSON APIs use dict/list/str/int/float/bool/None only at boundaries. Convert sets/tuples to lists for JSON.
Do not rely on dict order for semantic equality across systems. Compare as mappings, not key lists.
Freeze config dicts at startup when passing through layers. MappingProxyType or typed settings object.
Document empty container semantics - [] vs None. Callers should not guess missing vs empty.
Use Pydantic 2 models at HTTP boundaries; plain dicts inside hot domain logic when appropriate.

FAQs

When is OrderedDict still justified?

move_to_end LRU patterns. Regular dict handles insertion order otherwise.

list or tuple in API return?

Tuple for fixed small bundles; list when returning variable-length homogeneous collection.

Should I always use slots?

__slots__ for millions of small fixed-field objects. Premature for typical app records.

Counter vs defaultdict(int)?

Counter adds multiset operators and most_common - prefer for frequency work.

How avoid defaultdict serialization surprises?

Convert dict(defaultdict) before JSON - factory not serialized.

Are comprehensions always pythonic?

Yes for map/filter transforms. Switch to loop for side effects or complex branching.

dict vs dataclass for records?

dataclass/Pydantic for domain objects with behavior; dict for JSON passthrough or dynamic keys.

When use ChainMap?

Layered config search without copying - env over defaults. Flatten when persisting.

Is frozenset worth it?

When a set of tags must be a dict key or cache entry - otherwise plain set.

Biggest collections mistake?

Linear search on list inside loop - fix with set/dict index built once.

Choosing the Right Data Structure - decision guide
Data Structures Basics - intro examples
Immutability & Hashability - hash rules
collections Module - specialized types

The Hitchhiker's Guide to Python - Best-practices handbook for installation, configuration, style, and tooling.
Official Python Documentation - The definitive source (Tutorial, Library Reference, Language Reference, HOWTOs).
Python HOWTOs (Official) - Practical topic-focused guides.
Fluent Python (2nd Edition, O'Reilly) - Same book on O'Reilly.
Python Crash Course (3rd Edition) - Beginner-friendly projects (Eric Matthes).

Stack versions: This page was written for Python 3.14.0 (stable 3.14, maintenance 3.13), FastAPI 0.115+, Django 5.2, Flask 3.1, Pydantic 2, PyTorch 2.6+, pandas 2.2+, Polars 1.x, ruff 0.9+, and uv 0.6+.

Collections Best Practices

Practical rules for choosing and using Python's built-in containers without performance traps or readability debt.

How to Use This List

Apply during code review on data-heavy modules.
Revisit when profiling shows unexpected O(n) behavior.
Pair with type hints (dict[str, int]) for clearer contracts.
Use as onboarding checklist for data pipeline code.

A - Selection

Pick dict for keyed lookup, set for membership, list for ordered sequences. Document non-obvious choices.
Use deque for FIFO/LIFO queues instead of list.pop(0). O(1) vs O(n) at front.
Use heapq for priority scheduling; bisect for sorted insertions. Not sorted() on every insert.
Reach for pandas 2.2+ or Polars 1.x at tabular scale, not nested dicts of lists. Column ops and joins.
Prefer Counter and defaultdict over manual key-exists boilerplate. Fewer branches, clearer intent.

B - Correctness

Never use mutable lists or dicts as dict keys or set elements. Use tuple, frozenset, or frozen dataclass.
Shallow-copy nested structures with copy.deepcopy when independence required. dict.copy() shares inner lists.
Treat set iteration order as unspecified. Use sorted(s) for stable display only.
Validate uniqueness before inverting dict {v: k}. Colliding values overwrite silently.
Use get or try/except KeyError intentionally. d[k] when key must exist; get when optional.

C - Performance and Memory

Build set once for repeated membership in loops. Avoid O(n) in list in hot paths.
Use generator expressions feeding sum/any/max for large streams. Skip materializing million-item lists.
Consider array.array or bytes for compact numeric/binary buffers. Not list of ints for wire protocols.
Profile before swapping list for exotic structure. Small n favors simplicity.
Avoid repeated keys() snapshot unless mutating during iteration. list(d) when deleting while looping.

D - Readability

Use dict/set comprehensions for transforms with optional filters. Clearer than manual loops.
Name types in hints: dict[str, list[Event]], not bare dict. Self-documenting structures.
Use NamedTuple or dataclass instead of positional tuple indexing for records. row[3] obscures meaning.
Prefer | merge for immutable config overlays (3.9+). base | overrides reads left-to-right.
Order-preserving dedupe: list(dict.fromkeys(items)). Idiomatic one-liner.

E - Serialization and APIs

JSON APIs use dict/list/str/int/float/bool/None only at boundaries. Convert sets/tuples to lists for JSON.
Do not rely on dict order for semantic equality across systems. Compare as mappings, not key lists.
Freeze config dicts at startup when passing through layers. MappingProxyType or typed settings object.
Document empty container semantics - [] vs None. Callers should not guess missing vs empty.
Use Pydantic 2 models at HTTP boundaries; plain dicts inside hot domain logic when appropriate.

FAQs

When is OrderedDict still justified?

move_to_end LRU patterns. Regular dict handles insertion order otherwise.

list or tuple in API return?

Tuple for fixed small bundles; list when returning variable-length homogeneous collection.

Should I always use slots?

__slots__ for millions of small fixed-field objects. Premature for typical app records.

Counter vs defaultdict(int)?

Counter adds multiset operators and most_common - prefer for frequency work.

How avoid defaultdict serialization surprises?

Convert dict(defaultdict) before JSON - factory not serialized.

Are comprehensions always pythonic?

Yes for map/filter transforms. Switch to loop for side effects or complex branching.

dict vs dataclass for records?

dataclass/Pydantic for domain objects with behavior; dict for JSON passthrough or dynamic keys.

When use ChainMap?

Layered config search without copying - env over defaults. Flatten when persisting.

Is frozenset worth it?

When a set of tags must be a dict key or cache entry - otherwise plain set.

Biggest collections mistake?

Linear search on list inside loop - fix with set/dict index built once.

Choosing the Right Data Structure - decision guide
Data Structures Basics - intro examples
Immutability & Hashability - hash rules
collections Module - specialized types

The Hitchhiker's Guide to Python - Best-practices handbook for installation, configuration, style, and tooling.
Official Python Documentation - The definitive source (Tutorial, Library Reference, Language Reference, HOWTOs).
Python HOWTOs (Official) - Practical topic-focused guides.
Fluent Python (2nd Edition, O'Reilly) - Same book on O'Reilly.
Python Crash Course (3rd Edition) - Beginner-friendly projects (Eric Matthes).

Stack versions: This page was written for Python 3.14.0 (stable 3.14, maintenance 3.13), FastAPI 0.115+, Django 5.2, Flask 3.1, Pydantic 2, PyTorch 2.6+, pandas 2.2+, Polars 1.x, ruff 0.9+, and uv 0.6+.

Collections Best Practices

Search Documentation

Collections Best Practices