Skip to content

Course Map

Page Maps

graph LR
  family["Reproducible Research"]
  program["Deep Dive DVC"]
  section["Orientation and Study Practice"]
  page["Course Map"]
  journey["Beginner-to-stewardship journey"]

  family --> program --> section --> page
  page -.maps.-> journey
flowchart LR
  failures["failure diagnosis"] --> identity["data identity"]
  identity --> runtime["runtime influence"]
  runtime --> pipeline["stage contracts"]
  pipeline --> metrics["comparable meaning"]
  metrics --> experiments["controlled change"]
  experiments --> collaboration["shared execution"]
  collaboration --> recovery["survival"]
  recovery --> promotion["downstream trust"]
  promotion --> governance["long-lived ownership"]

The course progresses from one local question—“can this result be repeated?”—to a stewardship question—“which claims should this system continue to own?” Each module adds an evidence boundary needed by the next.

Use this map to plan complete study, locate a missing prerequisite, or explain why a later module cannot be learned as an isolated feature tutorial.

The four learning arcs

Arc Modules Governing question Evidence you learn to review
state truth 01–04 what state and influence produced this result? pointers, cache identity, runtime receipts, stage contracts
comparable change 05–06 do changed values describe the same claim and controlled experiment? metric contracts, parameters, baselines, experiment receipts
shared survival 07–08 can another executor reconstruct and recover the submitted state? CI evidence, remote state, recovery drills, retention decisions
downstream trust 09–10 what may consumers trust, and who should own each boundary? promotion manifests, integrity checks, migration and governance decisions

The arcs are not independent tracks. A metric comparison depends on truthful inputs; a promotion decision depends on comparable experiments; recovery depends on declared authority.

Modules 01–04: establish state truth

Module 01 — diagnose reproducibility claims

You learn to separate rerunning, repeating, reproducing, and reconstructing. The module starts from failure evidence because a tool choice cannot repair a claim that has not been defined.

Leave with:

  • an inventory of source, data, runtime, output, and operational evidence;
  • a repeatable-workflow claim tested against hidden influence;
  • language for rejecting claims broader than the evidence.

Module 02 — identify data by content

You trace the relationship among workspace paths, pointer files, cache objects, remotes, and published artifacts.

Leave able to:

  • explain why a path locates data but does not identify its bytes;
  • trace a pointer into cache and remote state;
  • predict which state move add, push, pull, or checkout performs;
  • test recovery from each surviving layer.

Module 03 — include execution environment influence

You treat runtime as a potential input surface rather than background scenery.

Leave able to:

  • separate graph evidence from runtime evidence;
  • choose lockfiles, containers, or CI by the boundary they control;
  • measure tolerance instead of claiming absolute determinism;
  • diagnose hidden runtime drift through causal contrast.

Module 04 — make pipelines truthful

You compare command reads, controls, and writes with DVC declarations.

Leave able to:

  • classify deps, selected params, and owned outputs;
  • predict status and reproduction from declared edges;
  • distinguish false reruns from stale results;
  • repair a hidden influence and replay the original mutation.
flowchart TD
  m1["M01: define the claim"] --> m2["M02: identify the bytes"]
  m2 --> m3["M03: identify runtime influence"]
  m3 --> m4["M04: declare causal edges"]
  m4 --> gate{"Can you explain why a result exists?"}
  gate -- no --> revisit["return to missing boundary"]
  gate -- yes --> compare["enter comparable change"]

Do not enter metric and experiment work merely because the pipeline executes. Enter when you can explain its state and influence boundaries.

Modules 05–06: establish comparable change

Module 05 — interpret metrics as semantic claims

A numeric difference is not automatically an improvement. You join population identity, metric definition, unit, aggregation, schema, and parameter controls.

Leave able to:

  • reconstruct the complete meaning behind a metric value;
  • distinguish mechanical DVC diffs from semantic comparison decisions;
  • accept additive schema evolution and reject incompatible drift;
  • abstain when comparability evidence is missing.

Module 06 — control experiments against an authority

Experiments become reviewable deviations from a named baseline rather than a collection of attractive results.

Leave able to:

  • identify baseline authority;
  • distinguish declared change from allowed change;
  • gate ranking on comparability;
  • separate candidate selection from promotion and cleanup.

Transition test:

Can you say which population, controls, metric contract, baseline, and allowed change make this candidate comparable?

If not, return to Modules 05–06 before treating a result as release evidence.

Modules 07–08: establish shared survival

Module 07 — make a reviewer handoff executable

You move from local success to submitted-state reconstruction by another executor.

Leave able to:

  • define shared remote stewardship;
  • preserve submitted-state CI evidence;
  • review merges as coherent data-and-code claims;
  • block handoffs that require oral knowledge or warm local caches.

Module 08 — prove recovery and retention

You distinguish current convenience from historical recoverability.

Leave able to:

  • define recovery objectives by protected revision;
  • reason about garbage-collection reachability;
  • migrate remotes without transferring authority prematurely;
  • preserve incident evidence and demonstrate cache-cold restoration.

Transition test:

Can a new executor recover the protected state using only governed declarations, metadata, remote objects, and documented credentials?

If not, downstream promotion rests on evidence the team may not be able to reconstruct.

Modules 09–10: establish downstream trust

Module 09 — create a smaller promotion boundary

You distinguish repository evidence from the artifact contract offered to consumers.

Leave able to:

  • bind promotion to an authorized baseline and policy;
  • build consumer-verifiable manifests and checksums;
  • detect threshold or evidence drift;
  • reject promotion when required integrity evidence is missing.

Module 10 — govern ownership and migration

You review whether DVC remains the right owner for data, execution, experiment, registry, release, and recovery concerns.

Leave able to:

  • name invariants before migration;
  • assign one authority to each state boundary;
  • design cutover, rollback, and acceptance evidence;
  • intervene on recurring contract failures rather than layering more tooling.

Suggested schedules

Schedule Rhythm Appropriate when
ten-week one technical module each week exercises and reflection need class discussion
five-week two adjacent modules each week learner already knows basic DVC commands
intensive one arc per study block learner can reserve long lab sessions
audit-led diagnostic route through modules by observed failure learner brings an existing repository

For accelerated study, do not drop exercises or evidence interpretation. Reduce repeated reading, not proof work.

Progress ledger

Track readiness by capability:

Boundary I can explain I can predict I can demonstrate I can review
data identity
runtime influence
stage contract
metric comparability
experiment control
reviewer handoff
recovery
promotion
governance

An empty cell is a navigation signal, not a score. Return to the earliest missing capability because later boundaries usually depend on it.

Route onward