Working in a large repository with sparse checkout and partial clone

What it is

A single repository with many projects is slow to clone and cluttered to work in. Git offers three independent reductions:

• Sparse checkout (git sparse-checkout set <dir>...) restricts which paths are materialised in the working tree and index. The documentation describes cone mode, now the default, where the input is a list of directories rather than gitignore-style patterns; the non-cone pattern mode is explicitly not recommended. --sparse-index shrinks the index to match.
• Partial clone (git clone --filter=blob:none) asks the server to omit objects according to a filter; blob:none omits all file contents until needed, blob:limit=<size> only blobs of at least that size. The design notes explain that the remote becomes a promisor remote and missing objects are fetched on demand, which requires being online and, because objects are fetched one at a time, "tends to be slow".
• Shallow clone (--depth <n>) truncates history. It is a different mechanism with its own limits and is not needed to make partial clone work.

Why it matters

Clone time, disk use and the size of git status scans grow with the whole repository, not with the part a person or a CI job touches. Applying the right reduction turns a multi-gigabyte checkout into a directory tree that fits the task.

How to apply

• For developers: git clone --filter=blob:none --sparse <url> (the --sparse option starts with only the top-level files), then git sparse-checkout set services/api libs/common.
• Keep whole directories in the cone; sibling files of every ancestor directory are included automatically, which is what makes build files at the root available.
• For CI jobs that need one commit: a blobless partial clone of the needed paths, or a shallow clone when history is irrelevant.
• Check git sparse-checkout list when a build cannot find a file; add the directory rather than disabling sparse mode.
• Commands that walk history (git log -p, git blame) trigger on-demand fetches in a partial clone; run them where the network is fast or prefetch first.

Pitfalls

Tools that scan the working tree assume the whole repository is present and may misreport missing directories. Non-cone patterns break --sparse-index and are slow. Offline work in a partial clone fails at the first missing blob. A shallow clone cannot answer any history question past its cutoff.