Installation

Using Pip

pip install parts

Or using the self-contained executable

./parts

Introduction

Parts is a build dependency and execution tool heavily inspired by SCons, bitbake, GitLab-CI among others.

The motivation for parts was the scale of a particular project being managed which had components larger and more independent than a typical SCons or Makefile targets; but not requireing the external package and configuration management provided by bitbake. These existing solution either felt awkward or overkill.

Parts intends to provide a simple, lightweight dependency tree API and runtime to build these medium sized projects. The command line interface and interation will feel very familiar.

Example

Assuming a folder structure

./parts.def
./module/parts.def

# parts.def
mod = include('module')

default(mod['library'])  # use the instance of Module created in module/parts.def

# module/parts.def
import asyncio
from parts import BuildStep


class Module(BuildStep):
    async def do_step(self):
        self.update_progress(total=100, phase='running')
        self.log.info(f'Running Module: {self.name}')
        for i in range(100):
            self.update_progress(advance=1, phase=f'loop {i:d}')
            await asyncio.sleep(0.1)  # do some work
        self.log.info(f'Completed Module: {self.name}')


Module('library')  # automatically registers target with name 'library'

Running

parts --no-cache

will result in

[11:29:56] INFO     Building ['default']...                                parts
           INFO     Running Module: library                              library
[11:30:06] INFO     Completed Module: library                            library
╭──────────────────────────────────────────────────────╮
│ 0:00:10 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1/1 │
╰──────────────────────────────────────────────────────╯

           INFO     Complete                                               parts

Design Goals

• parts.def use Python syntax with some shortcuts
• File paths are relative to the parts.def that references them
• BuildStep is the unit of work used by parts
• (sub)projects define all their BuildStep objects in a parts.def
• Build objects are useable by include()'ing another parts.def
• Build objects can be reproduced out of context if their input/output artefacts match (to enable distribution over multiple hosts)
• One parts.def can reference object instances created from an included parts.def
• BuildSteps can be used to communicate configuration options to depenent steps

Running parts

From within the toplevel of a project directory, next to a parts.def

parts [targets]

or by specifying the parts.def using --path

parts --path path/to/parts.def

Multiple targets can be specified simultaneously. If no target is provided then 'default' will be run, assuming it was provided by a call the default().

Listing available targets

parts --list

Graphing project dependencies

If graphviz is installed, parts can produce a graph for specific build targets to help visualize project dependencies.

parts --graph target

parts.def Reference

include()

# TODO

source()

# TODO

alias()

# TODO

copy()

# TODO

link()

# TODO

publish()

This will create a BuildStep that copies the given dependencies to the global ouput folder.

publish('a', 'b', BuildStep('c'))

will copy the outputs of task 'a', 'b' and the newly created 'c' to ./output.

alias('all', publish('a', 'b', 'c')

Will create a build target 'all' that publishes the outputs of 'a', 'b' and 'c'.

publish('a', dest='module_a')

Will publish ouputs of 'a' to the ./output/module_a folder.

arg()

# TODO

default()

# TODO

BuildStep

A unit of work in parts is represented by the abstract BuildStep which provides dependency linking with other BuildSteps as well as informative scheduling and execution of each task.

# TODO

Dependencies

# TODO

ProcessStep

# TODO

Elaborate vs Execution

Parts will elaborate all included config files to resolve their dependecies before beginning any execution.

This implies the inputs and outputs for each step must be defined during elaboration and will not change during execution.

While the execution of each step is parallelised (co-routines, though steps will likely produce their own processes), elaboration is serialized and not thread-safe.

Elaboration is fail-fast: the first config statement that raises aborts the whole run with the offending file:line, as do a syntax error and an include() of a path that doesn't exist. This stops the rest of a config running against a half-populated namespace and silently elaborating a different graph. --lenient restores the old skip-and-warn behaviour, which is useful for --listing a tree that is known to be broken.

Logging and Progress Bars

# TODO

ParseNodes

# TODO

State tracking

# TODO

phony targets

# TODO

parts.toml

Project-level configuration lives in parts.toml. All matching files are merged, with later locations overriding earlier ones section by section:

1. ~/.config/parts/runners.toml — legacy name, still accepted
2. ~/.config/parts/parts.toml — user-global
3. ./parts.runners.toml — legacy name, still accepted
4. ./parts.toml — project-local, takes precedence

So a user-global [runners.X] and a project-local [runners.Y] coexist, while a project-local [runners.X] overrides the user-global one.

The file currently has three sections:

• [parts] — overrides for the --build and --output CLI defaults. CLI flags still win when supplied.

  [parts]
  build  = "~/PyMoku-shared/build"
  output = "~/PyMoku-shared/output"
  

• [runners.<name>] — remote-runner rules (see below).

• [store] — the artifact store (see below).

Artifact store

A ProcessStep's cache has two tiers. The local one is the SQLite productions table next to the build (one row per output: rooted path, content hash, size, executable bit). The second is an artifact store speaking the Bazel Remote Execution API v2 — any conformant ActionCache + ContentAddressableStorage server (bazel-remote, NativeLink, BuildBarn), or a plain directory.

[store]
type = "disk"            # or "grpc"
path = "~/.cache/parts-store"
push = true              # false makes the store read-only for this checkout

[store]
type = "grpc"
url = "grpc://build.example.com:9092"   # grpcs:// for TLS
instance_name = "ben"
push = true
concurrency = 16          # transfers in flight while materialising

[store.headers]
authorization = "Bearer ${PARTS_STORE_TOKEN}"

With no [store] section nothing changes: the local tier is the whole cache.

What it stores

A step is keyed on its action digest (see Cache Invalidation). Against that key the store holds an ActionResult — every declared output as an output_files entry (rooted path, digest, executable bit), a creates('<pattern>') glob as an output_directories entry naming a Tree, and the run's stdout.log as stdout_digest — plus the output blobs themselves in the CAS.

Flows

• Read. On a local miss — or when the local rows describe outputs that are no longer on disk — parts asks the store for the ActionResult, writes it back as local rows, and treats the outputs as deliverable while FindMissingBlobs says the store still holds them. The step reports cached without running.
• Write. After a step produces its outputs, parts uploads the blobs the store is missing and updates the ActionResult — only when every declared output was hashed, and only when push is on.
• Deliver. Nothing is materialised until a reader that needs local bytes asks for it. Then the blob is downloaded, hash- and size-verified, into place, and its executable bit restored — the CAS stores bytes, the ActionResult stores the mode.

So deleting build/ and rebuilding re-materialises outputs from the store instead of re-running, and a second checkout with the same layout hits without running anything.

Who materialises what

Reader	Needs local bytes
plain BuildStep (Timing, BarFile, PythonZip…)	yes
ProcessStep on a local, Docker or ssh runner	yes
ProcessStep on a runner that takes inputs by digest (cloud)	no
copy()/link()	only if a reader of one of its outputs does
a requested target's deliverables	yes

A runner declares which it is with needs_local_inputs (True by default). A copy()/link() nobody local reads skips the copy entirely: its outputs already share their source's content hash, so they are recorded as usual and a later reader still gets them through the store — or, for a source the store never saw, by staging on demand.

Backends

• disk — ac/<hh>/<hash> holds a serialised ActionResult, cas/<hh>/<hash> holds blobs, written temp-and-rename. Uploads hardlink into the store when the filesystem allows it. A blob whose size doesn't match its digest reads as missing (so the step re-runs); one whose bytes don't match is refused on download. A shared filesystem makes this a usable team cache; it is not a wire format anyone else reads.
• grpc — a real REAPI client: ActionCache, ContentAddressableStorage and ByteStream. GetCapabilities supplies the batch size limit; blobs below it ride BatchUpdateBlobs/BatchReadBlobs, larger ones stream over ByteStream. instance_name scopes the keys; headers are sent as call metadata. Materialising many outputs reads small blobs 16 to a BatchReadBlobs and keeps concurrency requests in flight, shared across every step asking.

Custom types register through the parts.stores entry point group or register_store_type(), exactly like runners.

Remote Runners

By default, all ProcessStep commands run locally. Runners allow steps to be redirected to remote execution backends — SSH hosts, Docker containers, or cloud VMs — without changing parts.def files.

Each runner has a match pattern that selects which BuildStep classes it handles, and a type that determines the execution backend.

[runners.my-runner]
match = "VivadoStep|SynthStep"   # fnmatch patterns, pipe-separated
type = "ssh"                      # local, ssh, pool, docker, reapi, pymoku-cloud

Match patterns are tested against each class in the step's MRO (so matching VivadoStep also catches its subclasses). First matching rule wins. Unmatched steps run locally.

Runner Types

local

Runs commands via local subprocess. This is the default when no runner matches.

ssh

Executes on a remote host via SSH. Input files are rsynced to the remote host, the command runs there, and declared output files are rsynced back. Absolute paths in the command are rewritten to match the remote mirror.

[runners.build-server]
match = "VivadoStep"
type = "ssh"
host = "build.example.com"
user = "builder"
key = "~/.ssh/id_ed25519"
workdir = "~/parts-builds"
setup = "source /opt/tools/settings.sh"  # run before each command
jobs = 8                                  # max concurrent jobs

[runners.build-server.env]
LICENSE_FILE = "/opt/licenses/license.lic"

pool

Distributes jobs across multiple SSH hosts. Each host has a concurrency limit; jobs are dispatched to the first host with a free slot.

[runners.farm]
match = "VivadoStep"
type = "pool"
setup = "source /opt/tools/settings.sh"

[[runners.farm.hosts]]
host = "build1.example.com"
user = "builder"
key = "~/.ssh/id_ed25519"
workdir = "~/parts-builds"
jobs = 4

[[runners.farm.hosts]]
host = "build2.example.com"
user = "builder"
key = "~/.ssh/id_ed25519"
workdir = "~/parts-builds"
jobs = 8

docker

Runs commands inside a local Docker container.

[runners.sim]
match = "CocoTB*"
type = "docker"
image = "cocotb:latest"

reapi

Executes the step's REAPI Action on a remote-execution scheduler — the same objects the artifact store is keyed on, so nothing is packaged, tarred or polled.

[store]
type = "grpc"
url = "grpc://build.example.com:9092"

[runners.route]
match = "VivadoStep"
type = "reapi"
url = "grpc://build.example.com:9092"   # must be the store's server
instance_name = "ben"                   # defaults to the store's
priority = 10                           # ExecutionPolicy.priority
max_inflight = 200                      # concurrent Execute streams
exec_properties = { pool = "vivado-large" }

[runners.route.headers]
authorization = "Bearer ${PARTS_STORE_TOKEN}"

timeout, transfer_timeout and retries are accepted as for [store].

A [store] type = "grpc" on the same server is required: the scheduler reads the Action, the Command and every input by digest out of that CAS and publishes its outputs back into it, so the store is this runner's only artifact channel. A missing or differently-addressed store is a config error on the first step routed here.

The rule's exec_properties are routing properties — they are merged into an execution copy of the Action, never into the cache key, so which rule matched never affects cache hits. The step's own exec_properties() (what it needs of the platform: vivado = "2024.2", a container image) are in the key and reach the scheduler unchanged.

The runner takes its inputs by digest (needs_local_inputs = False), so a copy()/link() upstream of it never stages anything and an input a previous remote step produced is never downloaded. Only a source file that nothing has pushed is read off disk and uploaded before Execute.

Progress follows the Operation stream — ExecuteOperationMetadata.stage drives the phase and stdout_stream_name is read over ByteStream into the parse tree and <workdir>/stdout.log. The stream is reconnected with WaitExecution whenever it ends before the job does (a scheduler behind Cloud Run caps one request at 60 min; a route runs for hours).

pymoku-cloud

Submits each step as a job to an HTTP job API. Needs httpx (pip install pymoku-parts[cloud]) and a [store] on the same CAS the workers publish to.

[store]
type = "grpc"
url = "grpcs://store.example.com"

[runners.cloud]
match = "SynthStep|RouteStep"
type = "pymoku-cloud"
url = "https://jobs.example.com"
api_key = "${PARTS_API_KEY}"
worker_type = "large"                   # required: the worker pool
timeout = 28800                         # per-job wall clock, s (0: server default)
rule_priority = 10                      # claim order within a pool
max_inflight = 200                      # concurrent jobs per worker_type
max_jobs = 30                           # concurrent create/start requests
transient_patterns = ["licence checkout failed"]

Inputs go by digest: whatever the store lacks is uploaded before the job is created, and the worker fetches each one by hash. The worker publishes every output to the same CAS and the job's completion manifest names each one, so the outputs are recorded here and fetched only when something local reads them. A successful job whose manifest doesn't cover a declared output, or whose outputs the store doesn't hold, fails the step.

A failed job whose stdout.log contains one of transient_patterns is retried once on a fresh worker.

Custom runner types

External packages can register their own runner types via the parts.runners entry point group, or by calling register_runner_type() at import time.

Entry point (pyproject.toml):

[project.entry-points."parts.runners"]
myrunner = "mypackage.runner:create"

Where create(name, config) is a factory that returns a Runner instance.

Programmatic registration:

from pymoku.parts.runners import register_runner_type

def my_factory(name, config):
    return MyRunner(name, config)

register_runner_type('myrunner', my_factory)

Then use it in parts.toml:

[runners.example]
match = "SomeStep"
type = "myrunner"
custom_option = "value"

Cache Invalidation

A step's cache key (in_state) is SHA256(config_state + sorted input states). A ProcessStep instead keys on its REAPI action digest — SHA256(Action.SerializeToString(deterministic=True)) over {command_digest, input_root_digest, platform} — which covers the same things:

• Command.arguments = ['/bin/sh', '-c', cmd] with every configured root rewritten relative to the workdir, which is also the command the runner executes (cwd = the workdir, local or mirrored)
• Command.environment_variables — the declared environment (runner [env], then set_env(), then extra_env), sorted by name. Not os.environ: the local runner leaks that to the process, which is a hermeticity gap, not a cache input.
• Command.output_paths and working_directory, rooted
• Platform.properties from exec_properties() — what the step declares of its execution platform (a tool release, a container image id). Default {}; DockerStep reports container-image.
• the input root Directory tree: every input file at its rooted path (build/…, project/…, output/…; a path outside every root lands under abs/), with its content digest, size and executable bit.

Runner identity is not part of the key: a runner that provides a different tool than the step declares is a configuration error, not a different action. So switching runners does not invalidate anything, while editing a runner's [env] does.

Every run writes <workdir>/.action.json with the key and its components (in_state, action_digest, command_digest, input_root_digest, command, env, exec properties, input paths/hashes/sizes, output paths). Diff two of them to explain a cache miss.

Environment Variables

Environment variables are merged from three sources (later wins):

1. Runner config [runners.name.env] table
2. set_env() calls in parts.def (config_env)
3. ProcessStep extra_env kwarg

DockerStep with Remote Runners

When a DockerStep is matched by an SSH or cloud runner, the step's image attribute is forwarded. The remote runner wraps the command in docker run on the remote host, or passes it to the cloud API as a container job.