modal 0.68.11__py3-none-any.whl → 0.68.24__py3-none-any.whl

modal/__init__.py

@@ -17,6 +17,7 @@ try:
     from .cls import Cls, parameter
     from .dict import Dict
     from .exception import Error
+    from .file_pattern_matcher import FilePatternMatcher
     from .functions import Function
     from .image import Image
     from .mount import Mount
@@ -48,6 +49,7 @@ __all__ = [
     "Cron",
     "Dict",
     "Error",
+    "FilePatternMatcher",
     "Function",
     "Image",
     "Mount",

modal/_ipython.py

@@ -1,21 +1,11 @@
 # Copyright Modal Labs 2022
 import sys
-import warnings
-
-ipy_outstream = None
-try:
-    with warnings.catch_warnings():
-        warnings.simplefilter("ignore")
-        import ipykernel.iostream
-
-    ipy_outstream = ipykernel.iostream.OutStream
-except ImportError:
-    pass
 
 
 def is_notebook(stdout=None):
-    if ipy_outstream is None:
+    ipykernel_iostream = sys.modules.get("ipykernel.iostream")
+    if ipykernel_iostream is None:
         return False
     if stdout is None:
         stdout = sys.stdout
-    return isinstance(stdout, ipy_outstream)
+    return isinstance(stdout, ipykernel_iostream.OutStream)

modal/_runtime/asgi.py

@@ -1,4 +1,8 @@
 # Copyright Modal Labs 2022
+
+# Note: this module isn't imported unless it's needed.
+# This is because aiohttp is a pretty big dependency that adds significant latency when imported
+
 import asyncio
 from collections.abc import AsyncGenerator
 from typing import Any, Callable, NoReturn, Optional, cast

modal/_runtime/user_code_imports.py

@@ -9,15 +9,6 @@ import modal._runtime.container_io_manager
 import modal.cls
 import modal.object
 from modal import Function
-from modal._runtime.asgi import (
-    LifespanManager,
-    asgi_app_wrapper,
-    get_ip_address,
-    wait_for_web_server,
-    web_server_proxy,
-    webhook_asgi_app,
-    wsgi_app_wrapper,
-)
 from modal._utils.async_utils import synchronizer
 from modal._utils.function_utils import LocalFunctionError, is_async as get_is_async, is_global_object
 from modal.exception import ExecutionError, InvalidError
@@ -28,6 +19,7 @@ from modal_proto import api_pb2
 if typing.TYPE_CHECKING:
     import modal.app
     import modal.partial_function
+    from modal._runtime.asgi import LifespanManager
 
 
 @dataclass
@@ -36,7 +28,7 @@ class FinalizedFunction:
     is_async: bool
     is_generator: bool
     data_format: int  # api_pb2.DataFormat
-    lifespan_manager: Optional[LifespanManager] = None
+    lifespan_manager: Optional["LifespanManager"] = None
 
 
 class Service(metaclass=ABCMeta):
@@ -63,19 +55,22 @@ def construct_webhook_callable(
     webhook_config: api_pb2.WebhookConfig,
     container_io_manager: "modal._runtime.container_io_manager.ContainerIOManager",
 ):
+    # Note: aiohttp is a significant dependency of the `asgi` module, so we import it locally
+    from modal._runtime import asgi
+
     # For webhooks, the user function is used to construct an asgi app:
     if webhook_config.type == api_pb2.WEBHOOK_TYPE_ASGI_APP:
         # Function returns an asgi_app, which we can use as a callable.
-        return asgi_app_wrapper(user_defined_callable(), container_io_manager)
+        return asgi.asgi_app_wrapper(user_defined_callable(), container_io_manager)
 
     elif webhook_config.type == api_pb2.WEBHOOK_TYPE_WSGI_APP:
-        # Function returns an wsgi_app, which we can use as a callable.
-        return wsgi_app_wrapper(user_defined_callable(), container_io_manager)
+        # Function returns an wsgi_app, which we can use as a callable
+        return asgi.wsgi_app_wrapper(user_defined_callable(), container_io_manager)
 
     elif webhook_config.type == api_pb2.WEBHOOK_TYPE_FUNCTION:
         # Function is a webhook without an ASGI app. Create one for it.
-        return asgi_app_wrapper(
-            webhook_asgi_app(user_defined_callable, webhook_config.method, webhook_config.web_endpoint_docs),
+        return asgi.asgi_app_wrapper(
+            asgi.webhook_asgi_app(user_defined_callable, webhook_config.method, webhook_config.web_endpoint_docs),
             container_io_manager,
         )
 
@@ -86,11 +81,11 @@ def construct_webhook_callable(
         # We intentionally try to connect to the external interface instead of the loopback
         # interface here so users are forced to expose the server. This allows us to potentially
         # change the implementation to use an external bridge in the future.
-        host = get_ip_address(b"eth0")
+        host = asgi.get_ip_address(b"eth0")
         port = webhook_config.web_server_port
         startup_timeout = webhook_config.web_server_startup_timeout
-        wait_for_web_server(host, port, timeout=startup_timeout)
-        return asgi_app_wrapper(web_server_proxy(host, port), container_io_manager)
+        asgi.wait_for_web_server(host, port, timeout=startup_timeout)
+        return asgi.asgi_app_wrapper(asgi.web_server_proxy(host, port), container_io_manager)
     else:
         raise InvalidError(f"Unrecognized web endpoint type {webhook_config.type}")
 

modal/_utils/blob_utils.py

@@ -9,22 +9,22 @@ import time
 from collections.abc import AsyncIterator
 from contextlib import AbstractContextManager, contextmanager
 from pathlib import Path, PurePosixPath
-from typing import Any, BinaryIO, Callable, Optional, Union
+from typing import TYPE_CHECKING, Any, BinaryIO, Callable, Optional, Union
 from urllib.parse import urlparse
 
-from aiohttp import BytesIOPayload
-from aiohttp.abc import AbstractStreamWriter
-
 from modal_proto import api_pb2
 from modal_proto.modal_api_grpc import ModalClientModal
 
 from ..exception import ExecutionError
 from .async_utils import TaskContext, retry
 from .grpc_utils import retry_transient_errors
-from .hash_utils import UploadHashes, get_sha256_hex, get_upload_hashes
+from .hash_utils import UploadHashes, get_upload_hashes
 from .http_utils import ClientSessionRegistry
 from .logger import logger
 
+if TYPE_CHECKING:
+    from .bytes_io_segment_payload import BytesIOSegmentPayload
+
 # Max size for function inputs and outputs.
 MAX_OBJECT_SIZE_BYTES = 2 * 1024 * 1024  # 2 MiB
 
@@ -38,93 +38,16 @@ BLOB_MAX_PARALLELISM = 10
 # read ~16MiB chunks by default
 DEFAULT_SEGMENT_CHUNK_SIZE = 2**24
 
-
-class BytesIOSegmentPayload(BytesIOPayload):
-    """Modified bytes payload for concurrent sends of chunks from the same file.
-
-    Adds:
-    * read limit using remaining_bytes, in order to split files across streams
-    * larger read chunk (to prevent excessive read contention between parts)
-    * calculates an md5 for the segment
-
-    Feels like this should be in some standard lib...
-    """
-
-    def __init__(
-        self,
-        bytes_io: BinaryIO,  # should *not* be shared as IO position modification is not locked
-        segment_start: int,
-        segment_length: int,
-        chunk_size: int = DEFAULT_SEGMENT_CHUNK_SIZE,
-        progress_report_cb: Optional[Callable] = None,
-    ):
-        # not thread safe constructor!
-        super().__init__(bytes_io)
-        self.initial_seek_pos = bytes_io.tell()
-        self.segment_start = segment_start
-        self.segment_length = segment_length
-        # seek to start of file segment we are interested in, in order to make .size() evaluate correctly
-        self._value.seek(self.initial_seek_pos + segment_start)
-        assert self.segment_length <= super().size
-        self.chunk_size = chunk_size
-        self.progress_report_cb = progress_report_cb or (lambda *_, **__: None)
-        self.reset_state()
-
-    def reset_state(self):
-        self._md5_checksum = hashlib.md5()
-        self.num_bytes_read = 0
-        self._value.seek(self.initial_seek_pos)
-
-    @contextmanager
-    def reset_on_error(self):
-        try:
-            yield
-        except Exception as exc:
-            try:
-                self.progress_report_cb(reset=True)
-            except Exception as cb_exc:
-                raise cb_exc from exc
-            raise exc
-        finally:
-            self.reset_state()
-
-    @property
-    def size(self) -> int:
-        return self.segment_length
-
-    def md5_checksum(self):
-        return self._md5_checksum
-
-    async def write(self, writer: AbstractStreamWriter):
-        loop = asyncio.get_event_loop()
-
-        async def safe_read():
-            read_start = self.initial_seek_pos + self.segment_start + self.num_bytes_read
-            self._value.seek(read_start)
-            num_bytes = min(self.chunk_size, self.remaining_bytes())
-            chunk = await loop.run_in_executor(None, self._value.read, num_bytes)
-
-            await loop.run_in_executor(None, self._md5_checksum.update, chunk)
-            self.num_bytes_read += len(chunk)
-            return chunk
-
-        chunk = await safe_read()
-        while chunk and self.remaining_bytes() > 0:
-            await writer.write(chunk)
-            self.progress_report_cb(advance=len(chunk))
-            chunk = await safe_read()
-        if chunk:
-            await writer.write(chunk)
-            self.progress_report_cb(advance=len(chunk))
-
-    def remaining_bytes(self):
-        return self.segment_length - self.num_bytes_read
+# Files larger than this will be multipart uploaded. The server might request multipart upload for smaller files as
+# well, but the limit will never be raised.
+# TODO(dano): remove this once we stop requiring md5 for blobs
+MULTIPART_UPLOAD_THRESHOLD = 1024**3
 
 
 @retry(n_attempts=5, base_delay=0.5, timeout=None)
 async def _upload_to_s3_url(
     upload_url,
-    payload: BytesIOSegmentPayload,
+    payload: "BytesIOSegmentPayload",
     content_md5_b64: Optional[str] = None,
     content_type: Optional[str] = "application/octet-stream",  # set to None to force omission of ContentType header
 ) -> str:
@@ -180,6 +103,8 @@ async def perform_multipart_upload(
     upload_chunk_size: int = DEFAULT_SEGMENT_CHUNK_SIZE,
     progress_report_cb: Optional[Callable] = None,
 ) -> None:
+    from .bytes_io_segment_payload import BytesIOSegmentPayload
+
     upload_coros = []
     file_offset = 0
     num_bytes_left = content_length
@@ -273,6 +198,8 @@ async def _blob_upload(
             progress_report_cb=progress_report_cb,
         )
     else:
+        from .bytes_io_segment_payload import BytesIOSegmentPayload
+
         payload = BytesIOSegmentPayload(
             data, segment_start=0, segment_length=content_length, progress_report_cb=progress_report_cb
         )
@@ -305,9 +232,13 @@ async def blob_upload(payload: bytes, stub: ModalClientModal) -> str:
 
 
 async def blob_upload_file(
-    file_obj: BinaryIO, stub: ModalClientModal, progress_report_cb: Optional[Callable] = None
+    file_obj: BinaryIO,
+    stub: ModalClientModal,
+    progress_report_cb: Optional[Callable] = None,
+    sha256_hex: Optional[str] = None,
+    md5_hex: Optional[str] = None,
 ) -> str:
-    upload_hashes = get_upload_hashes(file_obj)
+    upload_hashes = get_upload_hashes(file_obj, sha256_hex=sha256_hex, md5_hex=md5_hex)
     return await _blob_upload(upload_hashes, file_obj, stub, progress_report_cb)
 
 
@@ -366,6 +297,7 @@ class FileUploadSpec:
     use_blob: bool
     content: Optional[bytes]  # typically None if using blob, required otherwise
     sha256_hex: str
+    md5_hex: str
     mode: int  # file permission bits (last 12 bits of st_mode)
     size: int
 
@@ -383,13 +315,15 @@ def _get_file_upload_spec(
         fp.seek(0)
 
         if size >= LARGE_FILE_LIMIT:
+            # TODO(dano): remove the placeholder md5 once we stop requiring md5 for blobs
+            md5_hex = "baadbaadbaadbaadbaadbaadbaadbaad" if size > MULTIPART_UPLOAD_THRESHOLD else None
             use_blob = True
             content = None
-            sha256_hex = get_sha256_hex(fp)
+            hashes = get_upload_hashes(fp, md5_hex=md5_hex)
         else:
             use_blob = False
             content = fp.read()
-            sha256_hex = get_sha256_hex(content)
+            hashes = get_upload_hashes(content)
 
     return FileUploadSpec(
         source=source,
@@ -397,7 +331,8 @@ def _get_file_upload_spec(
         mount_filename=mount_filename.as_posix(),
         use_blob=use_blob,
         content=content,
-        sha256_hex=sha256_hex,
+        sha256_hex=hashes.sha256_hex(),
+        md5_hex=hashes.md5_hex(),
         mode=mode & 0o7777,
         size=size,
     )

modal/_utils/bytes_io_segment_payload.py

@@ -0,0 +1,97 @@
+# Copyright Modal Labs 2024
+
+import asyncio
+import hashlib
+from contextlib import contextmanager
+from typing import BinaryIO, Callable, Optional
+
+# Note: this module needs to import aiohttp in global scope
+# This takes about 50ms and isn't needed in many cases for Modal execution
+# To avoid this, we import it in local scope when needed (blob_utils.py)
+from aiohttp import BytesIOPayload
+from aiohttp.abc import AbstractStreamWriter
+
+# read ~16MiB chunks by default
+DEFAULT_SEGMENT_CHUNK_SIZE = 2**24
+
+
+class BytesIOSegmentPayload(BytesIOPayload):
+    """Modified bytes payload for concurrent sends of chunks from the same file.
+
+    Adds:
+    * read limit using remaining_bytes, in order to split files across streams
+    * larger read chunk (to prevent excessive read contention between parts)
+    * calculates an md5 for the segment
+
+    Feels like this should be in some standard lib...
+    """
+
+    def __init__(
+        self,
+        bytes_io: BinaryIO,  # should *not* be shared as IO position modification is not locked
+        segment_start: int,
+        segment_length: int,
+        chunk_size: int = DEFAULT_SEGMENT_CHUNK_SIZE,
+        progress_report_cb: Optional[Callable] = None,
+    ):
+        # not thread safe constructor!
+        super().__init__(bytes_io)
+        self.initial_seek_pos = bytes_io.tell()
+        self.segment_start = segment_start
+        self.segment_length = segment_length
+        # seek to start of file segment we are interested in, in order to make .size() evaluate correctly
+        self._value.seek(self.initial_seek_pos + segment_start)
+        assert self.segment_length <= super().size
+        self.chunk_size = chunk_size
+        self.progress_report_cb = progress_report_cb or (lambda *_, **__: None)
+        self.reset_state()
+
+    def reset_state(self):
+        self._md5_checksum = hashlib.md5()
+        self.num_bytes_read = 0
+        self._value.seek(self.initial_seek_pos)
+
+    @contextmanager
+    def reset_on_error(self):
+        try:
+            yield
+        except Exception as exc:
+            try:
+                self.progress_report_cb(reset=True)
+            except Exception as cb_exc:
+                raise cb_exc from exc
+            raise exc
+        finally:
+            self.reset_state()
+
+    @property
+    def size(self) -> int:
+        return self.segment_length
+
+    def md5_checksum(self):
+        return self._md5_checksum
+
+    async def write(self, writer: "AbstractStreamWriter"):
+        loop = asyncio.get_event_loop()
+
+        async def safe_read():
+            read_start = self.initial_seek_pos + self.segment_start + self.num_bytes_read
+            self._value.seek(read_start)
+            num_bytes = min(self.chunk_size, self.remaining_bytes())
+            chunk = await loop.run_in_executor(None, self._value.read, num_bytes)
+
+            await loop.run_in_executor(None, self._md5_checksum.update, chunk)
+            self.num_bytes_read += len(chunk)
+            return chunk
+
+        chunk = await safe_read()
+        while chunk and self.remaining_bytes() > 0:
+            await writer.write(chunk)
+            self.progress_report_cb(advance=len(chunk))
+            chunk = await safe_read()
+        if chunk:
+            await writer.write(chunk)
+            self.progress_report_cb(advance=len(chunk))
+
+    def remaining_bytes(self):
+        return self.segment_length - self.num_bytes_read

modal/_utils/hash_utils.py

@@ -3,14 +3,14 @@ import base64
 import dataclasses
 import hashlib
 import time
-from typing import BinaryIO, Callable, Union
+from typing import BinaryIO, Callable, Optional, Sequence, Union
 
 from modal.config import logger
 
 HASH_CHUNK_SIZE = 65536
 
 
-def _update(hashers: list[Callable[[bytes], None]], data: Union[bytes, BinaryIO]) -> None:
+def _update(hashers: Sequence[Callable[[bytes], None]], data: Union[bytes, BinaryIO]) -> None:
     if isinstance(data, bytes):
         for hasher in hashers:
             hasher(data)
@@ -57,15 +57,44 @@ class UploadHashes:
     md5_base64: str
     sha256_base64: str
 
+    def md5_hex(self) -> str:
+        return base64.b64decode(self.md5_base64).hex()
 
-def get_upload_hashes(data: Union[bytes, BinaryIO]) -> UploadHashes:
+    def sha256_hex(self) -> str:
+        return base64.b64decode(self.sha256_base64).hex()
+
+
+def get_upload_hashes(
+    data: Union[bytes, BinaryIO], sha256_hex: Optional[str] = None, md5_hex: Optional[str] = None
+) -> UploadHashes:
     t0 = time.monotonic()
-    md5 = hashlib.md5()
-    sha256 = hashlib.sha256()
-    _update([md5.update, sha256.update], data)
+    hashers = {}
+
+    if not sha256_hex:
+        sha256 = hashlib.sha256()
+        hashers["sha256"] = sha256
+    if not md5_hex:
+        md5 = hashlib.md5()
+        hashers["md5"] = md5
+
+    if hashers:
+        updaters = [h.update for h in hashers.values()]
+        _update(updaters, data)
+
+    if sha256_hex:
+        sha256_base64 = base64.b64encode(bytes.fromhex(sha256_hex)).decode("ascii")
+    else:
+        sha256_base64 = base64.b64encode(hashers["sha256"].digest()).decode("ascii")
+
+    if md5_hex:
+        md5_base64 = base64.b64encode(bytes.fromhex(md5_hex)).decode("ascii")
+    else:
+        md5_base64 = base64.b64encode(hashers["md5"].digest()).decode("ascii")
+
     hashes = UploadHashes(
-        md5_base64=base64.b64encode(md5.digest()).decode("ascii"),
-        sha256_base64=base64.b64encode(sha256.digest()).decode("ascii"),
+        md5_base64=md5_base64,
+        sha256_base64=sha256_base64,
     )
-    logger.debug("get_upload_hashes took %.3fs", time.monotonic() - t0)
+
+    logger.debug("get_upload_hashes took %.3fs (%s)", time.monotonic() - t0, hashers.keys())
     return hashes

modal/_utils/http_utils.py

@@ -1,18 +1,18 @@
 # Copyright Modal Labs 2022
 import contextlib
-import socket
-import ssl
-from typing import Optional
+from typing import TYPE_CHECKING, Optional
 
-import certifi
-from aiohttp import ClientSession, ClientTimeout, TCPConnector
-from aiohttp.web import Application
-from aiohttp.web_runner import AppRunner, SockSite
+# Note: importing aiohttp seems to take about 100ms, and it's not really necessarily,
+# unless we need to work with blobs. So that's why we import it lazily instead.
+
+if TYPE_CHECKING:
+    from aiohttp import ClientSession
+    from aiohttp.web import Application
 
 from .async_utils import on_shutdown
 
 
-def _http_client_with_tls(timeout: Optional[float]) -> ClientSession:
+def _http_client_with_tls(timeout: Optional[float]) -> "ClientSession":
     """Create a new HTTP client session with standard, bundled TLS certificates.
 
     This is necessary to prevent client issues on some system where Python does
@@ -22,13 +22,18 @@ def _http_client_with_tls(timeout: Optional[float]) -> ClientSession:
     Specifically: the error "unable to get local issuer certificate" when making
     an aiohttp request.
     """
+    import ssl
+
+    import certifi
+    from aiohttp import ClientSession, ClientTimeout, TCPConnector
+
     ssl_context = ssl.create_default_context(cafile=certifi.where())
     connector = TCPConnector(ssl=ssl_context)
     return ClientSession(connector=connector, timeout=ClientTimeout(total=timeout))
 
 
 class ClientSessionRegistry:
-    _client_session: ClientSession
+    _client_session: "ClientSession"
     _client_session_active: bool = False
 
     @staticmethod
@@ -47,9 +52,13 @@ class ClientSessionRegistry:
 
 
 @contextlib.asynccontextmanager
-async def run_temporary_http_server(app: Application):
+async def run_temporary_http_server(app: "Application"):
     # Allocates a random port, runs a server in a context manager
     # This is used in various tests
+    import socket
+
+    from aiohttp.web_runner import AppRunner, SockSite
+
     sock = socket.socket()
     sock.bind(("", 0))
     port = sock.getsockname()[1]

modal/_utils/{pattern_matcher.py → pattern_utils.py}

@@ -5,7 +5,7 @@ This is the same pattern-matching logic used by Docker, except it is written in
 Python rather than Go. Also, the original Go library has a couple deprecated
 functions that we don't implement in this port.
 
-The main way to use this library is by constructing a `PatternMatcher` object,
+The main way to use this library is by constructing a `FilePatternMatcher` object,
 then asking it whether file paths match any of its patterns.
 """
 
@@ -148,75 +148,6 @@ class Pattern:
             return False
 
 
-class PatternMatcher:
-    """Allows checking paths against a list of patterns."""
-
-    def __init__(self, patterns: list[str]) -> None:
-        """Initialize a new PatternMatcher instance.
-
-        Args:
-            patterns (list): A list of pattern strings.
-
-        Raises:
-            ValueError: If an illegal exclusion pattern is provided.
-        """
-        self.patterns: list[Pattern] = []
-        self.exclusions = False
-        for pattern in patterns:
-            pattern = pattern.strip()
-            if not pattern:
-                continue
-            pattern = os.path.normpath(pattern)
-            new_pattern = Pattern()
-            if pattern[0] == "!":
-                if len(pattern) == 1:
-                    raise ValueError('Illegal exclusion pattern: "!"')
-                new_pattern.exclusion = True
-                pattern = pattern[1:]
-                self.exclusions = True
-            # In Python, we can proceed without explicit syntax checking
-            new_pattern.cleaned_pattern = pattern
-            new_pattern.dirs = pattern.split(os.path.sep)
-            self.patterns.append(new_pattern)
-
-    def matches(self, file_path: str) -> bool:
-        """Check if the file path or any of its parent directories match the patterns.
-
-        This is equivalent to `MatchesOrParentMatches()` in the original Go
-        library. The reason is that `Matches()` in the original library is
-        deprecated due to buggy behavior.
-        """
-        matched = False
-        file_path = os.path.normpath(file_path)
-        if file_path == ".":
-            # Don't let them exclude everything; kind of silly.
-            return False
-        parent_path = os.path.dirname(file_path)
-        if parent_path == "":
-            parent_path = "."
-        parent_path_dirs = parent_path.split(os.path.sep)
-
-        for pattern in self.patterns:
-            # Skip evaluation based on current match status and pattern exclusion
-            if pattern.exclusion != matched:
-                continue
-
-            match = pattern.match(file_path)
-
-            if not match and parent_path != ".":
-                # Check if the pattern matches any of the parent directories
-                for i in range(len(parent_path_dirs)):
-                    dir_path = os.path.sep.join(parent_path_dirs[: i + 1])
-                    if pattern.match(dir_path):
-                        match = True
-                        break
-
-            if match:
-                matched = not pattern.exclusion
-
-        return matched
-
-
 def read_ignorefile(reader: TextIO) -> list[str]:
     """Read an ignore file from a reader and return the list of file patterns to
     ignore, applying the following rules:

modal/_utils/shell_utils.py

@@ -19,14 +19,20 @@ def write_to_fd(fd: int, data: bytes):
     future = loop.create_future()
 
     def try_write():
+        nonlocal data
         try:
             nbytes = os.write(fd, data)
-            loop.remove_writer(fd)
-            future.set_result(nbytes)
+            data = data[nbytes:]
+            if not data:
+                loop.remove_writer(fd)
+                future.set_result(None)
         except OSError as e:
-            if e.errno != errno.EAGAIN:
-                future.set_exception(e)
-                raise
+            if e.errno == errno.EAGAIN:
+                # Wait for the next write notification
+                return
+            # Fail if it's not EAGAIN
+            loop.remove_writer(fd)
+            future.set_exception(e)
 
     loop.add_writer(fd, try_write)
     return future