Recipes by workload

The features section is organised by what a thing is. This section is organised by what you are doing, because the right API usually follows from the workload rather than from the operation.

Four workloads

Interactive request/response →One input, an answer now. Priorities: latency, predictability, code you can read at 3 a.m.Streaming →Input larger than memory, or output needed before input ends. Priorities: bounded memory, lazy processing, early output.Batch corpora →Many documents, offline, throughput is the metric. Priorities: memory reuse, shared setup, allocation removal.Massive parallel corpora →Enough work to be worth threads. Priorities: CPU utilisation, chunk sizing, per-worker state.

The workload decides the API

The same operation, four times:

Workload	Tokenizer call	What you optimise	Watch out for
Interactive	tokenize()	Latency and readability	Doing O(n) work over a whole corpus per request
Streaming	tokens()	Peak memory, time to first result	Any collect() in the middle of the pipeline
Batch	tokenize_into()	Allocations per document	Forgetting buf.clear() at the top of the loop
Parallel	map_init(Vec::new, …) + tokenize_into()	CPU utilisation	Tasks too small to pay for scheduling

None of those rows is "the fast one" — they are answers to different questions. See Choosing the right API.

Problem recipes

Complete programs for common tasks:

Fuzzy name matching →Phonetic bucketing to cut the candidate set, then edit distance to rank. The order matters more than the metric.Prefix autocomplete →A trie, a lazy prefix iterator, and why keys_with_prefix is the wrong call when you only show ten suggestions.

Before you optimise

Every recipe here that uses a performance-oriented API also says what it costs you in code complexity, because that is a real cost. Two rules of thumb:

• Optimise the sequential version first. A tokenize_into loop that removes ten million allocations may make threads unnecessary — and it composes with them if not.
• Measure before and after. Batch is the one workload where these techniques reliably show up. If the numbers do not move, put the simple version back.

If you are not in the workload a recipe was written for, take the simpler version from Your first program instead.