Offload literals

flytekit/bin/entrypoint.py

@@ -14,7 +14,7 @@
 import uuid
 import warnings
 from sys import exit
-from typing import Callable, List, Optional
+from typing import Callable, Dict, List, Optional
 
 import click
 from flyteidl.core import literals_pb2 as _literals_pb2
@@ -55,6 +55,7 @@
 from flytekit.models.core import identifier as _identifier
 from flytekit.tools.fast_registration import download_distribution as _download_distribution
 from flytekit.tools.module_loader import load_object_from_module
+from flytekit.utils.pbhash import compute_hash_string
 
 
 def get_version_message():
@@ -135,6 +136,7 @@ def _dispatch_execute(
     load_task: Callable[[], PythonTask],
     inputs_path: str,
     output_prefix: str,
+    is_map_task: bool = False,
 ):
     """
     Dispatches execute to PythonTask
@@ -144,6 +146,12 @@ def _dispatch_execute(
             a: [Optional] Record outputs to output_prefix
             b: OR if IgnoreOutputs is raised, then ignore uploading outputs
             c: OR if an unhandled exception is retrieved - record it as an errors.pb
+
+    :param ctx: FlyteContext
+    :param load_task: Callable[[], PythonTask]
+    :param inputs: Where to read inputs
+    :param output_prefix: Where to write primitive outputs
+    :param is_map_task: Whether this task is executing as part of a map task
     """
     error_file_name = _build_error_file_name()
     worker_name = _get_worker_name()
@@ -179,7 +187,59 @@ def _dispatch_execute(
             logger.warning("Task produces no outputs")
             output_file_dict = {_constants.OUTPUT_FILE_NAME: _literal_models.LiteralMap(literals={})}
         elif isinstance(outputs, _literal_models.LiteralMap):
-            output_file_dict = {_constants.OUTPUT_FILE_NAME: outputs}
+            # The keys in this map hold the filenames to the offloaded proto literals.
+            offloaded_literals: Dict[str, _literal_models.Literal] = {}
+            literal_map_copy = {}
+
+            offloading_enabled = os.environ.get("_F_L_MIN_SIZE_MB", None) is not None
+            min_offloaded_size = -1
+            max_offloaded_size = -1
+            if offloading_enabled:
+                min_offloaded_size = int(os.environ.get("_F_L_MIN_SIZE_MB", "10")) * 1024 * 1024
+                max_offloaded_size = int(os.environ.get("_F_L_MAX_SIZE_MB", "1000")) * 1024 * 1024
+
+            # Go over each output and create a separate offloaded in case its size is too large
+            for k, v in outputs.literals.items():
+                literal_map_copy[k] = v
+
+                if not offloading_enabled:
+                    continue
+
+                lit = v.to_flyte_idl()
+                if max_offloaded_size != -1 and lit.ByteSize() >= max_offloaded_size:
+                    raise ValueError(
+                        f"Literal {k} is too large to be offloaded. Max literal size is {max_offloaded_size} whereas the literal size is {lit.ByteSize()} bytes"
+                    )
+
+                if min_offloaded_size != -1 and lit.ByteSize() >= min_offloaded_size:
+                    logger.debug(f"Literal {k} is too large to be inlined, offloading to metadata bucket")
+                    inferred_type = task_def.interface.outputs[k].type
+
+                    # In the case of map tasks we need to use the type of the collection as inferred type as the task
+                    # typed interface of the offloaded literal. This is done because the map task interface present in
+                    # the task template contains the (correct) type for the entire map task, not the single node execution.
+                    # For that reason we "unwrap" the collection type and use it as the inferred type of the offloaded literal.
+                    if is_map_task:
+                        inferred_type = inferred_type.collection_type
+
+                    # This file will hold the offloaded literal and will be written to the output prefix
+                    # alongside the regular outputs.pb, deck.pb, etc.
+                    # N.B.: by construction `offloaded_filename` is guaranteed to be unique
+                    offloaded_filename = f"{k}_offloaded_metadata.pb"
+                    offloaded_literal = _literal_models.Literal(
+                        offloaded_metadata=_literal_models.LiteralOffloadedMetadata(
+                            uri=f"{output_prefix}/{offloaded_filename}",
+                            size_bytes=lit.ByteSize(),
+                            # TODO: remove after https://github.com/flyteorg/flyte/pull/5909 is merged
+                            inferred_type=inferred_type,
+                        ),
+                        hash=v.hash if v.hash is not None else compute_hash_string(lit),
+                    )
+                    literal_map_copy[k] = offloaded_literal
+                    offloaded_literals[offloaded_filename] = v
+            outputs = _literal_models.LiteralMap(literals=literal_map_copy)
+
+            output_file_dict = {_constants.OUTPUT_FILE_NAME: outputs, **offloaded_literals}
         elif isinstance(outputs, _dynamic_job.DynamicJobSpec):
             output_file_dict = {_constants.FUTURES_FILE_NAME: outputs}
         else:
@@ -588,7 +648,7 @@ def load_task():
             )
             return
 
-        _dispatch_execute(ctx, load_task, inputs, output_prefix)
+        _dispatch_execute(ctx, load_task, inputs, output_prefix, is_map_task=True)
 
 
 def normalize_inputs(

flytekit/core/task.py

@@ -5,12 +5,7 @@
 from functools import update_wrapper
 from typing import Any, Callable, Dict, Iterable, List, Optional, Tuple, Type, TypeVar, Union, overload
 
-from flytekit.core.utils import str2bool
-
-try:
-    from typing import ParamSpec
-except ImportError:
-    from typing_extensions import ParamSpec  # type: ignore
+from typing_extensions import ParamSpec  # type: ignore
 
 from flytekit.core import launch_plan as _annotated_launchplan
 from flytekit.core import workflow as _annotated_workflow
@@ -20,6 +15,7 @@
 from flytekit.core.python_function_task import PythonFunctionTask
 from flytekit.core.reference_entity import ReferenceEntity, TaskReference
 from flytekit.core.resources import Resources
+from flytekit.core.utils import str2bool
 from flytekit.deck import DeckField
 from flytekit.extras.accelerators import BaseAccelerator
 from flytekit.image_spec.image_spec import ImageSpec

flytekit/core/workflow.py

@@ -8,13 +8,9 @@
 from functools import update_wrapper
 from typing import Any, Callable, Coroutine, Dict, List, Optional, Tuple, Type, Union, cast, overload
 
+from typing_extensions import ParamSpec  # type: ignore
 from typing_inspect import is_optional_type
 
-try:
-    from typing import ParamSpec
-except ImportError:
-    from typing_extensions import ParamSpec  # type: ignore
-
 from flytekit.core import constants as _common_constants
 from flytekit.core import launch_plan as _annotated_launch_plan
 from flytekit.core.base_task import PythonTask, Task

flytekit/interaction/string_literals.py

@@ -61,6 +61,9 @@ def literal_string_repr(lit: Literal) -> typing.Any:
         return [literal_string_repr(i) for i in lit.collection.literals]
     if lit.map:
         return {k: literal_string_repr(v) for k, v in lit.map.literals.items()}
+    if lit.offloaded_metadata:
+        # TODO: load literal from offloaded literal?
+        return f"Offloaded literal metadata: {lit.offloaded_metadata}"
     raise ValueError(f"Unknown literal type {lit}")
 
 

flytekit/models/literals.py

@@ -991,7 +991,7 @@ def to_flyte_idl(self):
             map=self.map.to_flyte_idl() if self.map is not None else None,
             hash=self.hash,
             metadata=self.metadata,
-            offloaded_metadata=self.offloaded_metadata.to_flyte_idl() if self.offloaded_metadata else None,
+            offloaded_metadata=self.offloaded_metadata.to_flyte_idl() if self.offloaded_metadata is not None else None,
         )
 
     @classmethod

flytekit/utils/pbhash.py

@@ -0,0 +1,39 @@
+# This is a module that provides hashing utilities for Protobuf objects.
+import base64
+import hashlib
+import json
+
+from google.protobuf import json_format
+from google.protobuf.message import Message
+
+
+def compute_hash(pb: Message) -> bytes:
+    """
+    Computes a deterministic hash in bytes for the Protobuf object.
+    """
+    try:
+        pb_dict = json_format.MessageToDict(pb)
+        # json.dumps with sorted keys to ensure stability
+        stable_json_str = json.dumps(
+            pb_dict, sort_keys=True, separators=(",", ":")
+        )  # separators to ensure no extra spaces
+    except Exception as e:
+        raise ValueError(f"Failed to marshal Protobuf object {pb} to JSON with error: {e}")
+
+    try:
+        # Deterministically hash the JSON object to a byte array. Using SHA-256 for hashing here,
+        # assuming it provides a consistent hash output.
+        hash_obj = hashlib.sha256(stable_json_str.encode("utf-8"))
+    except Exception as e:
+        raise ValueError(f"Failed to hash JSON for Protobuf object {pb} with error: {e}")
+
+    # The digest is guaranteed to be 32 bytes long
+    return hash_obj.digest()
+
+
+def compute_hash_string(pb: Message) -> str:
+    """
+    Computes a deterministic hash in base64 encoded string for the Protobuf object
+    """
+    hash_bytes = compute_hash(pb)
+    return base64.b64encode(hash_bytes).decode("utf-8")