User Guide: Review vs AuditPhabricator User Documentation (Application User Guides)
Discusses the differences between "review" and "audit" workflows.
Phabricator supports two similar but separate code review workflows: "review" and "audit".
Review occurs in Differential, before changes are published. You can learn more in Differential User Guide.
Audit occurs in Diffusion, after changes are published. You can learn more in Audit User Guide.
When this documentation discusses "unpublished changes", it refers to changes which are still subject to being reworked in response to feedback. In many workflows, these changes will only exist locally on the developer's machine, but some workflows push tentative or temporary changes into remotes. The step that "publishes" changes might be either pushing or merging them, depending on your workflow.
Both the audit and review workflows are lightweight, asynchronous web-based workflows where reviewers or auditors inspect code independently, from their own machines -- not synchronous review sessions where authors and reviewers meet in person to discuss changes.
Broadly, review is normally a blocking workflow: in review workflows, authors usually can not publish changes until review completes and reviewers are satisfied.
In contrast, audit is normally a nonblocking workflow: in audit workflows, changes usually move forward by default.
Pre-publish review is significantly more powerful than post-publish auditing. You gain these advantages by requiring review before changes may be published:
- Authors have a strong incentive to craft small, well-formed changes that will be readily understood, to explain them adequately, and to provide appropriate test plans, test coverage and context.
- Reviewers have a real opportunity to make significant suggestions about architecture or approach in review. These suggestions are less attractive to adopt from audit, and may be much more difficult to adopt if significant time has passed between publish and audit.
- Authors have a strong incentive to fix problems and respond to feedback received during review because it blocks them. Authors have a much weaker incentive to promptly address problems raised during audit.
- Authors can ask reviewers to apply and verify fixes before they are published.
- Authors can easily pursue feedback early, and get course corrections on approach or direction.
- Reviewers are better prepared to support a given change once it is in production, having already had a chance to become familiar with and reason through the code.
- Reviewers are able to catch problems which automated tests may have difficulty detecting. For example, human reviewers are able to reason about performance problems that tests can easily miss because they run on small datasets and stub out service calls.
- Communicating about changes before they happen generally leads to better preparation for their effects.
The theoretical cost of review is that it slows down development by introducing a blocking step into the process and generally wastes developer time that could be better spent developing. This is less true than it appears, because the costs are low and pay for themselves in other ways:
- Differential is fast and provides a lightweight process for submitting code for review and for performing review.
- Authors are free to pursue other changes while code is being reviewed. With appropriate change management (like local branching in Git) they can even pursue dependent changes easily. Authors should rarely if ever be blocked on review, even though an individual change is blocked until it is approved.
- The workflow as a whole is lightweight and, with skillful reviewers, effective at identifying bugs. It is generally faster to fix bugs in review than in production.
- More importantly, it is effective at identifying problems with architecture and approach. These are free to fix in review ("don't do this, it is a bad idea") and may be very time consuming to fix in production. No matter how good your test suite is, it can't identify solutions which are poor because of missing context, or miscommunication, or which are simply bad ideas.
- Changes which are too large or too complicated to be reviewed quickly are often too large and too complicated, period. Nearly all large changes can be split apart into small, independent pieces which are easier to understand and test. Review tends to encourage smaller and better-factored changes.
- Review can be integrated with static analysis which can detect (and, in many cases, correct) mechanical problems with code like syntax, formatting, naming conventions, style problems, misspellings, and some program errors. This reduces the amount of time it takes to review code, and means reviewers can focus on actual problems with the code rather than minor stylistic issues.
- Review creates a permanent record of context and intent which explains why a change was made, generally with much more information than commit messages alone (authors have an incentive to properly explain a change when sending it for review). This makes it easier to understand code later, and to respond appropriately when it breaks.
- With arc patch, it is roughly as easy to pull a change out of Differential as it is to pull it out of the remote.
Post-publish audit is a less powerful workflow than pre-publish review, but can supplement review and is better than nothing on its own. If you are unpersuaded by the arguments above (or work on a team that is unswayed), audits provide some of the benefits of review with less friction:
- Audits are driven entirely by Phabricator: users do not need to install arc.
- Audits require little adjustment to existing workflows and little training.
- Audits are completely nonblocking, and send fewer notifications than review.
- Even if you have review, audits can be useful as a supplement to keep tabs on lower-importance changes or raise issues that are discovered after review.
Here are super biased recommendations from developers of code review software:
- If you can do review, do it. Supplement it with audits for less important changes as your organization scales.
- If you can't do review immediately, set up audits and try to transition toward review. Some types of changes (like tentative changes or requests for feedback about code) are a naturally good fit for review and can serve as a stepping stone toward broader acceptance. Greater familiarity with the toolset may also foster more acceptance toward review, and the value of review may become more obvious as the organization scales (e.g., once you get interns).
- If you aren't interested in review, just do audits. You can always change your mind later. But consider review! It's really good, we promise!