1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
|
"""Turn a `container.toml` into a `modal.Image` and a `modal.App`.
Every container -- under `containers/` here, `.modal/` in a repo that merely
builds itself on Modal -- is a directory with a `container.toml` and
a stub `container.py`. The keys are documented by the comments below, which
is the whole of the spec. Nothing here is Modal-specific configuration in its
own right -- each
spec key maps onto a documented Modal argument, and the mapping is meant to
stay boring enough to read straight through.
"""
import os
import subprocess
import tomllib
import modal
class SpecError(Exception):
"""The container.toml says something that cannot be built."""
def _running_in_modal() -> bool:
"""True inside a Modal container, false on the machine that launched it."""
return bool(os.environ.get("MODAL_TASK_ID"))
class Container:
"""One container: its spec, its image, its app, and how to run it."""
def __init__(self, spec: dict, directory: str):
self.is_remote = _running_in_modal()
self.dir = directory
self.spec = spec
self.name = self._require("container", "name")
self.description = spec.get("container", {}).get("description", "")
run = spec.get("run", {})
self.workdir = run.get("workdir", "/app")
self.command = run.get("command", "")
self.env = dict(run.get("env", {}))
# "function" (the default) or "sandbox". Sandboxes can run on a real
# VM, which Functions cannot -- see ../README.md.
self.runtime = spec.get("container", {}).get("runtime", "function")
# name -> mount path. Modal Volumes, mounted while the container runs
# and NOT while its image is built: a volume mount is not part of the
# resulting image, so anything written to one during a build step is
# gone by the time the container starts. Persist across runs is the
# whole point -- a toolchain too big to fetch every run, a build
# tree an incremental compile reuses, a dataset too big to bake in.
self.volume_spec = dict(spec.get("volumes", {}))
# Ports to tunnel out of a Sandbox, from [network] ports. Encrypted,
# which is Modal's own word for "the tunnel terminates TLS and speaks
# plain HTTP to your process" -- so the thing listening inside is an
# ordinary http.server and not something holding a certificate.
#
# Sandbox-only: a Function has no long-lived process to tunnel into,
# and `modal run` on one would hand back a URL for a container that
# has already exited.
self.ports = [int(p) for p in spec.get("network", {}).get("ports", [])]
# Modal re-imports this module inside the container, so everything
# below runs twice: once here, once out there. Out there the local
# tree does not exist -- no repo to copy from -- and
# the image is already built, so validating and rebuilding it would
# only fail. The App still has to exist for the decorators to bind.
if self.is_remote:
self.image = None
self.app = modal.App(self.name)
else:
self._validate()
self.image = self._build_image()
self.app = modal.App(self.name, image=self.image)
# -- spec reading ----------------------------------------------------
@classmethod
def from_toml(cls, container_py: str) -> "Container":
"""Load the container.toml sitting next to the given container.py."""
directory = os.path.dirname(os.path.abspath(container_py))
path = os.path.join(directory, "container.toml")
if not os.path.exists(path):
raise SpecError(f"no container.toml in {directory}")
with open(path, "rb") as f:
return cls(tomllib.load(f), directory)
def _require(self, table: str, key: str):
try:
return self.spec[table][key]
except KeyError:
raise SpecError(f"container.toml needs [{table}] {key}") from None
def _validate(self):
c = self.spec.get("container", {})
if self.runtime not in ("function", "sandbox"):
raise SpecError(
f'[container] runtime must be "function" or "sandbox",'
f" not {self.runtime!r}"
)
if bool(c.get("base")) == bool(c.get("registry")):
raise SpecError("set exactly one of [container] base or registry")
if self.ports and self.runtime != "sandbox":
raise SpecError(
"[network] ports needs [container] runtime = \"sandbox\" --"
" a Function has no process to tunnel into"
)
for name, mount in self.volume_spec.items():
if not isinstance(mount, str) or not mount.startswith("/"):
raise SpecError(
f"[volumes] {name} must be an absolute path, not {mount!r}"
)
# Mounting over one of these hides what the image already has
# there: an empty volume would shadow what the image's own build
# put in its place.
if mount.rstrip("/") in ("", "/usr", "/etc", "/bin", "/lib"):
raise SpecError(
f"[volumes] {name} may not mount over {mount} --"
" it would hide what the image has there"
)
if mount.rstrip("/") == self.workdir.rstrip("/"):
raise SpecError(
f"[volumes] {name} may not mount over the workdir"
f" ({mount}) -- [build] include copies land there"
)
# -- image -----------------------------------------------------------
def _build_image(self) -> modal.Image:
c = self.spec["container"]
build = self.spec.get("build", {})
if c.get("base"):
image = modal.Image.from_name(c["base"])
else:
image = modal.Image.from_registry(c["registry"])
# copy=True throughout: later run_commands need these files present.
# `context` is what include paths are relative to, and it may sit above
# the container directory -- a container that builds the repo it lives
# in sets context = "../..", so include = ["."] means the whole repo.
context = os.path.normpath(os.path.join(self.dir, build.get("context", ".")))
# Image building and program building, kept apart.
#
# `[build] commands` run AFTER the source copy, so any edit anywhere
# in the tree invalidates them. `warm` + `setup` are the same step
# moved in front of the source: `warm` names only the files the step
# actually reads, and `setup` runs against those alone, so editing
# `flutter/src` invalidates nothing above the final copy. A container
# whose setup is an `apt-get` needs no `warm` at all -- it reads
# nothing out of the tree, so nothing in the tree can invalidate it.
for rel in build.get("warm", []):
src = os.path.normpath(os.path.join(context, rel))
dest = f"{self.workdir}/{rel}"
if os.path.isdir(src):
image = image.add_local_dir(src, dest, copy=True)
else:
image = image.add_local_file(src, dest, copy=True)
if setup := build.get("setup", []):
image = image.run_commands(*setup, volumes=self.volumes)
# `ignore` is what keeps a build tree out of the image. A checkout that
# has been built in locally carries its output -- flutter/build and the
# caches beside it were 395MB of a 441MB repo -- and all of it would be
# uploaded on every start only to be thrown away, since the container
# builds into a volume of its own. Patterns are relative to the copied
# directory, as in .dockerignore.
ignore = list(build.get("ignore", []))
for rel in build.get("include", ["."]):
src = os.path.normpath(os.path.join(context, rel))
dest = self.workdir if rel == "." else f"{self.workdir}/{rel}"
if os.path.isdir(src):
image = image.add_local_dir(src, dest, copy=True, ignore=ignore)
else:
image = image.add_local_file(src, dest, copy=True)
# A repo copied in brings its `.git` along, and in a worktree that is
# a *file* holding `gitdir: <path on the machine that copied it>` --
# which points at nothing out here, so any tool that follows it fails
# in a way that has nothing to do with what it was asked to do.
# Nothing in a container wants the git metadata, so it goes.
image = image.run_commands(f"rm -rf {self.workdir}/.git")
if commands := build.get("commands", []):
# Volumes mounted for the build too, not just the run, so a step
# can read a cache the last run filled. The mount is not part of
# the resulting image; only what the step writes outside it is.
image = image.run_commands(*commands, volumes=self.volumes)
# container.py does `from _loader import Container`, and Modal mounts
# the entrypoint file alone -- so without this the import that works
# locally fails in the container. copy=False adds it at startup rather
# than baking a layer, so it invalidates nothing above it, and it must
# therefore come after every build step.
image = image.add_local_python_source("_loader")
# ...and container.py reads its spec at import time, so the spec has to
# be there too. Modal re-imports the entrypoint at /root, which is the
# one directory that gets none of the workdir copies above.
image = image.add_local_file(
os.path.join(self.dir, "container.toml"), "/root/container.toml"
)
return image
# -- volumes ---------------------------------------------------------
@property
def volumes(self) -> dict:
"""{mount path: Volume}, as both Sandbox.create and @app.function want.
`from_name` is lazy, so this is safe to evaluate on the re-import
inside the container as well as out here. create_if_missing means a
spec naming a volume that does not exist yet makes it rather than
failing -- the first run of a cache is the one that fills it.
"""
return {
mount: modal.Volume.from_name(name, create_if_missing=True)
for name, mount in self.volume_spec.items()
}
# -- function --------------------------------------------------------
@property
def function_kwargs(self) -> dict:
"""Everything `@app.function` should be given, from [resources]."""
r = self.spec.get("resources", {})
kwargs: dict = {"timeout": int(r.get("timeout", 900))}
if r.get("cpu"):
kwargs["cpu"] = float(r["cpu"])
if r.get("memory"):
kwargs["memory"] = int(r["memory"])
if r.get("gpu"):
kwargs["gpu"] = r["gpu"]
# Deliberately NOT self.experimental_options: that fills in the
# sandbox default, and vm_runtime on a Function is refused by the
# server. A sandbox container's @app.function is vestigial anyway.
if self.runtime == "function":
if experimental := dict(self.spec.get("experimental", {})):
kwargs["experimental_options"] = experimental
if volumes := self.volumes:
kwargs["volumes"] = volumes
return kwargs
@property
def experimental_options(self) -> dict:
"""Experimental options, with the sandbox default filled in.
`vm_runtime` is Sandbox-only: the server rejects it on a Function
outright. So it is defaulted on for sandboxes and never for functions,
and an explicit [experimental] table always wins.
"""
explicit = dict(self.spec.get("experimental", {}))
if explicit:
return explicit
if self.runtime == "sandbox":
return {"vm_runtime": True}
return {}
@property
def sandbox_kwargs(self) -> dict:
"""Everything `Sandbox.create` should be given, from [resources]."""
r = self.spec.get("resources", {})
kwargs: dict = {"timeout": int(r.get("timeout", 900))}
if r.get("cpu"):
kwargs["cpu"] = float(r["cpu"])
if r.get("memory"):
# A VM sandbox gets exactly this much and cannot grow into more.
kwargs["memory"] = int(r["memory"])
if opts := self.experimental_options:
kwargs["experimental_options"] = dict(opts)
if volumes := self.volumes:
kwargs["volumes"] = volumes
if self.ports:
kwargs["encrypted_ports"] = list(self.ports)
return kwargs
def shell_command(self, override: str = "") -> str:
"""The full shell line the container runs."""
command = override or self.command
if not command:
raise SpecError("container.toml has no [run] command")
return f"cd {self.workdir} && {command}"
def run_sandbox(self, override: str = "") -> str:
"""Run one command in a Sandbox that dies when the command does.
The command IS the sandbox's process, rather than something exec'd
into a `sleep infinity` box that then has to be torn down. There is no
idle window to pay for and nothing to leak if this script is killed;
the [resources] timeout is a backstop, not the mechanism.
Modal streams a Sandbox's output into the app log as it runs -- which
is what you want for a build -- so this returns "" rather than handing
back a copy for the caller to print underneath it.
"""
line = self.shell_command(override)
sb = modal.Sandbox.create(
"sh", "-c", line,
app=self.app,
image=self.image,
workdir=self.workdir,
env={k: str(v) for k, v in self.env.items()},
**self.sandbox_kwargs,
)
self._print_tunnels(sb)
sb.wait()
if sb.returncode != 0:
raise RuntimeError(
f"{self.name}: command failed ({sb.returncode})\n"
f"$ {line}\n{sb.stderr.read()}"
)
return ""
def _print_tunnels(self, sb: "modal.Sandbox") -> None:
"""Say where a tunnelled port can be reached, once the sandbox is up.
`tunnels()` blocks until the Sandbox is scheduled, which is why this
is called after create and not folded into it. Nothing prints when
[network] ports is empty, which is every container but the serving
ones.
"""
if not self.ports:
return
for port, tunnel in sb.tunnels().items():
print(f" :{port} -> {tunnel.url}")
def open_sandbox(self) -> "modal.Sandbox":
"""Start a Sandbox and leave it running, for `scripts/shell`.
Same workdir and the same [run] env as the real thing: a shell opened
to debug a container that does not have the container's environment is
a shell that reproduces something else. `command` is the one part left
out, because not running it is the point.
"""
return modal.Sandbox.create(
"sleep",
"infinity",
app=self.app,
image=self.image,
workdir=self.workdir,
env={k: str(v) for k, v in self.env.items()},
**self.sandbox_kwargs,
)
def execute(self, override: str = "") -> str:
"""Run the command in this container. Called remotely, not locally."""
line = self.shell_command(override)
result = subprocess.run(
line,
shell=True,
capture_output=True,
text=True,
env={**os.environ, **{k: str(v) for k, v in self.env.items()}},
)
if result.returncode != 0:
raise RuntimeError(
f"{self.name}: command failed ({result.returncode})\n"
f"$ {line}\n{result.stderr}"
)
return result.stdout
|