我知道这个问题已经有了一些流行的答案。但是有一种为包管理器缓存文件的新方法。我认为在未来，当BuildKit变得更加标准时，这可能是一个很好的答案。
从Docker 18.09开始，就有了对BuildKit的实验性支持。BuildKit增加了对Dockerfile中一些新特性的支持，包括将外部卷挂载到RUN步骤的实验性支持。这允许我们为$HOME/.cache/pip/这样的东西创建缓存。
我们将以下面的requirements.txt文件为例:
Click==7.0
Django==2.2.3
django-appconf==1.0.3
django-compressor==2.3
django-debug-toolbar==2.0
django-filter==2.2.0
django-reversion==3.0.4
django-rq==2.1.0
pytz==2019.1
rcssmin==1.0.6
redis==3.3.4
rjsmin==1.1.0
rq==1.1.0
six==1.12.0
sqlparse==0.3.0
一个典型的Python Dockerfile示例如下:
FROM python:3.7
WORKDIR /usr/src/app
COPY requirements.txt /usr/src/app/
RUN pip install -r requirements.txt
COPY . /usr/src/app
使用DOCKER_BUILDKIT环境变量启用BuildKit，我们可以在大约65秒内构建非缓存pip步骤:
$ export DOCKER_BUILDKIT=1
$ docker build -t test .
[+] Building 65.6s (10/10) FINISHED
=> [internal] load .dockerignore 0.0s
=> => transferring context: 2B 0.0s
=> [internal] load build definition from Dockerfile 0.0s
=> => transferring dockerfile: 120B 0.0s
=> [internal] load metadata for docker.io/library/python:3.7 0.5s
=> CACHED [1/4] FROM docker.io/library/python:3.7@sha256:6eaf19442c358afc24834a6b17a3728a45c129de7703d8583392a138ecbdb092 0.0s
=> [internal] load build context 0.6s
=> => transferring context: 899.99kB 0.6s
=> CACHED [internal] helper image for file operations 0.0s
=> [2/4] COPY requirements.txt /usr/src/app/ 0.5s
=> [3/4] RUN pip install -r requirements.txt 61.3s
=> [4/4] COPY . /usr/src/app 1.3s
=> exporting to image 1.2s
=> => exporting layers 1.2s
=> => writing image sha256:d66a2720e81530029bf1c2cb98fb3aee0cffc2f4ea2aa2a0760a30fb718d7f83 0.0s
=> => naming to docker.io/library/test 0.0s
现在，让我们添加实验头文件，并修改RUN步骤来缓存Python包:
# syntax=docker/dockerfile:experimental
FROM python:3.7
WORKDIR /usr/src/app
COPY requirements.txt /usr/src/app/
RUN --mount=type=cache,target=/root/.cache/pip pip install -r requirements.txt
COPY . /usr/src/app
现在继续执行另一个构建。应该花同样多的时间。但这一次它是在我们的新缓存挂载中缓存Python包:
$ docker build -t pythontest .
[+] Building 60.3s (14/14) FINISHED
=> [internal] load build definition from Dockerfile 0.0s
=> => transferring dockerfile: 120B 0.0s
=> [internal] load .dockerignore 0.0s
=> => transferring context: 2B 0.0s
=> resolve image config for docker.io/docker/dockerfile:experimental 0.5s
=> CACHED docker-image://docker.io/docker/dockerfile:experimental@sha256:9022e911101f01b2854c7a4b2c77f524b998891941da55208e71c0335e6e82c3 0.0s
=> [internal] load .dockerignore 0.0s
=> [internal] load build definition from Dockerfile 0.0s
=> => transferring dockerfile: 120B 0.0s
=> [internal] load metadata for docker.io/library/python:3.7 0.5s
=> CACHED [1/4] FROM docker.io/library/python:3.7@sha256:6eaf19442c358afc24834a6b17a3728a45c129de7703d8583392a138ecbdb092 0.0s
=> [internal] load build context 0.7s
=> => transferring context: 899.99kB 0.6s
=> CACHED [internal] helper image for file operations 0.0s
=> [2/4] COPY requirements.txt /usr/src/app/ 0.6s
=> [3/4] RUN --mount=type=cache,target=/root/.cache/pip pip install -r requirements.txt 53.3s
=> [4/4] COPY . /usr/src/app 2.6s
=> exporting to image 1.2s
=> => exporting layers 1.2s
=> => writing image sha256:0b035548712c1c9e1c80d4a86169c5c1f9e94437e124ea09e90aea82f45c2afc 0.0s
=> => naming to docker.io/library/test 0.0s
大约60秒。类似于我们的第一个构建。
对requirements.txt做一个小的修改(比如在两个包之间增加一个新行)来强制缓存无效并再次运行:
$ docker build -t pythontest .
[+] Building 15.9s (14/14) FINISHED
=> [internal] load build definition from Dockerfile 0.0s
=> => transferring dockerfile: 120B 0.0s
=> [internal] load .dockerignore 0.0s
=> => transferring context: 2B 0.0s
=> resolve image config for docker.io/docker/dockerfile:experimental 1.1s
=> CACHED docker-image://docker.io/docker/dockerfile:experimental@sha256:9022e911101f01b2854c7a4b2c77f524b998891941da55208e71c0335e6e82c3 0.0s
=> [internal] load build definition from Dockerfile 0.0s
=> => transferring dockerfile: 120B 0.0s
=> [internal] load .dockerignore 0.0s
=> [internal] load metadata for docker.io/library/python:3.7 0.5s
=> CACHED [1/4] FROM docker.io/library/python:3.7@sha256:6eaf19442c358afc24834a6b17a3728a45c129de7703d8583392a138ecbdb092 0.0s
=> CACHED [internal] helper image for file operations 0.0s
=> [internal] load build context 0.7s
=> => transferring context: 899.99kB 0.7s
=> [2/4] COPY requirements.txt /usr/src/app/ 0.6s
=> [3/4] RUN --mount=type=cache,target=/root/.cache/pip pip install -r requirements.txt 8.8s
=> [4/4] COPY . /usr/src/app 2.1s
=> exporting to image 1.1s
=> => exporting layers 1.1s
=> => writing image sha256:fc84cd45482a70e8de48bfd6489e5421532c2dd02aaa3e1e49a290a3dfb9df7c 0.0s
=> => naming to docker.io/library/test 0.0s
只有大约16秒!
我们得到这个加速是因为我们不再下载所有的Python包。它们由包管理器(在本例中为pip)缓存，并存储在缓存卷挂载中。卷挂载提供给运行步骤，以便pip可以重用我们已经下载的包。这发生在任何Docker层缓存之外。
在更大的requirements.txt上，增益应该会更好。
注:
这是实验性的Dockerfile语法，应该被这样对待。您目前可能不想在生产环境中使用它进行构建。
BuildKit目前还不能在Docker Compose或其他直接使用Docker API的工具下工作。Docker Compose在1.25.0就支持这个功能了。参见如何使用docker-compose启用BuildKit ?
目前还没有用于托管缓存的直接接口。当你做一个docker系统修剪-a时，它被清除。
希望这些特性能够加入到Docker的构建中，而BuildKit将成为默认配置。如果/当这种情况发生时，我会尝试更新这个答案。
