Companion labs: examples you can check | How Data Works

The supplied practice package uses Python’s standard library and fresh synthetic data. It does not connect to your website, payment provider, production database or cloud account. Read this guide before running code.

Start in an isolated folder#

Extract practice-code.zip into a new folder. Keep its Python files together because the later suites import the earlier teaching modules.
Use Python 3.11 or newer with SQLite 3.37.0 or newer. This minimum covers the syntax and STRICT tables used here; it is not a recommendation to operate an old runtime in production. The accompanying test-results.json identifies the actual authoring environment.
Open a terminal in that extracted folder and run the test command below. It discovers the six test modules. Some cases deliberately confirm a missing safeguard or a failing design; their success means the stated counterexample was observed.
The examples include in-memory databases and one earlier bounded temporary-file locking demonstration. Temporary resources are cleaned up by their tests. No personal database or user records are required.

python3 -m unittest discover -s . -p 'test_*.py' -v
python3 advanced_lab.py
python3 capstone_lab.py

The last two commands print a small local performance observation and a capstone transaction/replay/restore trace. Timing results vary by machine and workload. They do not promise that a particular index always wins.

What is implemented#

Module	Chapters and exercises	Boundary
`foundations_lab.py`	1–3: canonical orders, number/text/time representations and NULL behavior	Pure functions and a fresh SQLite fixture.
`history_lab.py`	4–6: measurements, scoped identity, duplicates and reconstructed history	Small explicit state machines; not a production event store.
`data_bridge_lab.py`	7–9 and 13: CSV/JSON import, the shop tables, relationships, totals, backup and a lock example	Bounded input profile, synthetic fixtures and local SQLite.
`schema_sql_lab.py`	10–12: schema rules, SQL, parameters, rollback and deliberately missing protections	Product-specific SQLite observations; PostgreSQL is documented, not executed.
`advanced_lab.py`	14–46: query measurements, a leaf split, hashes, wait graphs, revisions, LRU, recovery prefixes, tombstones, replicas, fencing, majorities, protocol states, caches, windows, manifests, encodings, search, vectors and uncertainty	Local models and isolated observations, not full storage engines or network protocols.
`capstone_lab.py`	47–52: permission fixtures, tenant-scoped orders, outbox/inbox tasks, retries, local restore, retention status, resumable backfill and capacity/cost arithmetic	Trusted Principal objects, in-memory SQLite and bounded functions; not authentication, a public API or a deployed service.

The advanced models#

The B-tree exercise implements ordered separator selection and a single overflowing-leaf split. It does not implement a complete balanced persistent B-tree. The log example replays committed toy transactions from a supplied durable prefix; it does not parse an engine’s actual WAL or simulate torn storage sectors. The compaction example can discard tombstones only under its stated complete-history, latest-only assumptions.

The replica model tracks contiguous applied positions. The fencing model has the resource check a monotonically increasing authority token. The consensus exercises enumerate majorities, compare last-term/last-index pairs and test the current-term direct-commit condition. These are selected rules, not a full Raft implementation or proof of progress under a real network.

The transaction participant and compensation classes show explicit state transitions and repeated-decision handling. The cache and manifest classes model generation/version checks in one process; they do not implement shared-memory atomicity or a distributed metadata service.

The stream examples distinguish fixed windows, sliding windows, session overlap and late corrections. The columnar exercise packs signed 64-bit integers and uses zlib to compare representations. It is not a Parquet writer or a disk-I/O benchmark. Text search uses a tiny tokenized corpus, with an additional SQLite FTS5 check when that feature is present. Vector examples calculate exact distances and filter by tenant before selecting results; they do not train an embedding model or implement a production approximate-nearest-neighbor index.

The Wilson interval, missing-status bounds and nearest-rank percentiles check arithmetic under explicit assumptions. They do not establish representative sampling, independence or causal identification in real business records.

The capstone boundary#

The capstone accepts a canonicalized cart, reads current catalogue prices during acceptance, conditionally reduces stock and stores the agreed order, request identity and outbox event in one owned transaction. An identical scoped replay returns the committed result without repricing or reducing stock again. A conflicting payload using the same scoped request identity is rejected.

The consumer stores one local task and its inbox identity atomically. An injected lost acknowledgement occurs after that local consumer commit; redelivery does not create another task. This is not proof that an external courier, email provider or payment processor performs its work exactly once.

The local restore uses SQLite’s backup API, enables foreign-key checking on the target, checks database integrity and references, compares logical records and continues a synthetic operation. It does not test power loss, cloud loss, independent failure domains or an off-site recovery objective.

The Principal objects are trusted fixtures constructed by the test. They are not login tokens. Holding the raw SQLite connection bypasses the helper authorization boundary; a test explicitly demonstrates this limit. The model does not enforce immutable agreements against an administrator or implement a complete checkout lifecycle.

Reading the test record#

The release test record gives the exact test count, failures, errors, skips, runtime versions and hashes of the executable files. A skipped feature check is not a pass for that feature. Test names and assertions are the evidence; the number of tests is only a count.

Exercises elsewhere in the book can ask you to draw a schedule, inspect a plan, choose a workload or design a real restore procedure. Not every such task is a ready-made automated implementation. PostgreSQL deployments, actual network faults, device flush behavior, complete tenant security and live schema migrations remain tasks for an appropriately isolated environment with explicit authorization.