@creative-introvert/prediction-testing
v0.3.0
Published
`prediction-testing` is a test runner, simlar to [Jest](https://jestjs.io/), [vitest](https://vitest.dev/) or [ava](https://github.com/avajs/ava), that focuses on testing predictive functions (e.g. search, auto-complete, or ML).
Downloads
26
Readme
prediction-testing
prediction-testing
is a test runner, simlar to Jest, vitest or ava, that focuses on testing predictive functions (e.g. search, auto-complete, or ML).
When you SHOULD use prediction-testing
- you have tons of test cases, and your tests purely compare inputs and outputs of the function under test
- you are testing a statistical model, or a predictive function, where 100% successful test results are impossible or impractical
- you are testing a (flaky) legacy system
When you SHOULD NOT use prediction-testing
- when your tests are few, and predominantely example-based; use the conventional test runners instead (jest, ava, vitest)
- when you require your testing framework to do all sorts of magic (auto-mocking, spies, etc)
Before Getting Started
Before looking at code examples, some notes on the design:
- Unlike tools like jest or vitest, prediction-testing's CLI does not provide a runtime, but is imported as a library. Check the "Why No Runtime?" section on the reasoning.
- Though only required minimally, the library depends on using effect (the missing standard library for TypeScript). At minimum, your function under test has to return an Effect. If your function doesn't already do so, checkout the Usage below, to see how to trivially convert it.
Further, a cautionary note on stability:
- The library is still in alpha, and will, regularly, without warning, release breaking changes.
- A beta phase is planned, which will keep releasing breaking changes, but provide guidance on migration paths, or even code-mods for automatic migration.
Usage
With your package manager of choice, install the following packages:
@creative-introvert/prediction-testing
@creative-introvert/prediction-testing-cli
effect
The examples use pnpm
and pnpx
+ tsx
to execute the files. Replace with
whatever you use in your setup (e.g. npm
and ts-node
, or yarn
and tsc +
node
, etc.)
With CLI
Define your test-suite.
// my-test-suite.ts
import * as CLI from '@creative-introvert/prediction-testing-cli';
import {Effect} from 'effect';
const myFunction = (input: number) => Promise.resolve(input * 1.7);
void CLI.run({
testSuite: {
// Convert myFunction to Effect-returning.
program: (input: number) => Effect.promise(() => myFunction(input)),
testCases: [
{input: 0, expected: 0},
{input: 1, expected: 2},
{input: 2, expected: 3},
{input: 3, expected: 4},
{input: 4, expected: 5},
],
},
testSuiteName: 'simple',
// Currently, test results are written to the file-system.
// This will be replaced by an SQLite backend soon™.
dirPath: '.metrics',
filePostfix: 'ptest',
});
Summarize
pnpx tsx my-test-suite.ts summarize
=============================================================================
SUMMARY
=============================================================================
index | id | input | tags | label | expected | result | previous result
-----------------------------------------------------------------------------
1/5 │ 5feceb66 │ 0 │ │ TP │ 0 │ 0 │ ∅
-----------------------------------------------------------------------------
2/5 │ 6b86b273 │ 1 │ │ FP │ 2 │ 1.7 │ ∅
-----------------------------------------------------------------------------
3/5 │ d4735e3a │ 2 │ │ FP │ 3 │ 3.4 │ ∅
-----------------------------------------------------------------------------
4/5 │ 4e074085 │ 3 │ │ FP │ 4 │ 5.1 │ ∅
-----------------------------------------------------------------------------
5/5 │ 4b227777 │ 4 │ │ FP │ 5 │ 6.8 │ ∅
=============================================================================
index | id | input | tags | label | expected | result | previous result
=============================================================================
======================================
STATS
======================================
TP | TN | FP | FN | precision | recall
--------------------------------------
1 | 0 | 4 | 0 | 0.20 | 1.00
======================================
Summarize With Labels Filter
pnpx tsx my-test-suite.ts summarize --labels TP
=============================================================================
SUMMARY
=============================================================================
index | id | input | tags | label | expected | result | previous result
-----------------------------------------------------------------------------
1/1 │ 5feceb66 │ 0 │ │ TP │ 0 │ 0 │ ∅
=============================================================================
index | id | input | tags | label | expected | result | previous result
=============================================================================
======================================
STATS
======================================
TP | TN | FP | FN | precision | recall
--------------------------------------
1 | 0 | 0 | 0 | 1.00 | 1.00
======================================
Write Test Results
pnpx tsx my-test-suite.ts write
Diff
Assuming you have (1) previous test results, and (2) something changed, e.g. the inputs/expecations of your test suite, or the function-under-test implementation.
diff --git a/my-test-suite.ts b/my-test-suite.ts
index 21cd713..ab1b6dc 100644
--- a/my-test-suite.ts
+++ b/my-test-suite.ts
@@ -1,7 +1,7 @@
import * as CLI from '@creative-introvert/prediction-testing-cli';
import {Effect} from 'effect';
-const myFunction = (input: number) => Promise.resolve(input * 1.7);
+const myFunction = (input: number) => Promise.resolve(input * 2);
void CLI.run({
testSuite: {
pnpx tsx my-test-suite.ts diff
# or, if you want it to process.exit(1) on diff
pnpx tsx my-test-suite.ts diff --ci
=============================================================================
SUMMARY
=============================================================================
index | id | input | tags | label | expected | result | previous result
-----------------------------------------------------------------------------
1/4 │ 6b86b273 │ 1 │ │ TP │ 2 │ 2 │ 1.7
-----------------------------------------------------------------------------
2/4 │ d4735e3a │ 2 │ │ FP │ 3 │ 4 │ 3.4
-----------------------------------------------------------------------------
3/4 │ 4e074085 │ 3 │ │ FP │ 4 │ 6 │ 5.1
-----------------------------------------------------------------------------
4/4 │ 4b227777 │ 4 │ │ FP │ 5 │ 8 │ 6.8
=============================================================================
index | id | input | tags | label | expected | result | previous result
=============================================================================
======================================
DIFF
======================================
TP | TN | FP | FN | precision | recall
--------------------------------------
0 │ 0 │ -1 │ 0 │ 0.05 │ 0.00
======================================
Checkout workspace/examples/src/with-cli
for more examples.
pnpx tsx <file-path>
# e.g.
pnpx tsx ./workspace/examples/src/with-cli/simple.ts
Without CLI
For full control over execution, and if you're not afraid of using effect you may simply import the runner functions individually.
import * as PT from '@creative-introvert/prediction-testing';
import {Console, Duration, Effect, Stream} from 'effect';
type Input = number;
type Result = number;
const program = (input: Input) =>
Effect.promise(() => Promise.resolve(input * 2)).pipe(
Effect.delay(Duration.millis(500)),
);
const testCases: PT.TestCase<Input, Result>[] = [
{input: 0, expected: 0},
{input: 1, expected: 1},
{input: 2, expected: 2},
{input: 3, expected: 3},
{input: 4, expected: 4},
];
await PT.testAll({
testCases,
program,
}).pipe(
PT.runFoldEffect,
Effect.tap(testRun => Console.log(PT.Show.summary({testRun}))),
Effect.runPromise,
);
Checkout workspace/examples/src/as-library
for more examples.
pnpx tsx <file-path>
# e.g.
pnpx tsx ./workspace/examples/src/as-library/simple.ts
Why No Runtime?
Tools like jest and vitest provide a dedicated CLI, which the user can run,
separately from other entrypoints into their app. For example, jest provides
the jest
binary, which, magically, finds all the test files, and executes
them, reporting on the results.
Unfortunately, this comes with a lot of complexity:
- Unless the user provides their tests as ES5 JavaScript (not a thing these days), jest has to figure out how transpile/compile the source. Given the level of, let's call it, variation in JavaScript land (Typescript, and its many different configurations, custom runtimes like svelte, ESM vs CommonJS, etc), this is no easy feat. The code required for enabling this would easily overshadow the actual test runner in terms of complexity.
- As
jest
(etc) essentially behave as a framework (jest calls your code) as opposed to a library (you call jest), customization requires additional complexity in the framework, which now has to provide various entrypoints into its execution.
Alternatively, the library could hook into an existing test runner (I've seen that vite provides some programatic context), but I have not yet looked deeper into that. Though saving me from solving this problem myself, this would likely come with its own limitations.
Providing prediction-testing
as a library is maybe not as satisfying and
convenient as a dedicated binary, but is vastly simpler, and far more
powerful, enabling trivial extensibility for the user.