feat: add min_concurrency

fal-ai · Jan 21, 2024 · 6ccf6ba · 6ccf6ba
1 parent 9b42407
commit 6ccf6ba
Show file tree

Hide file tree

Showing 6 changed files with 109 additions and 46 deletions.
diff --git a/projects/fal/src/fal/api.py b/projects/fal/src/fal/api.py
@@ -32,6 +32,7 @@
 from fal.sdk import (
     FAL_SERVERLESS_DEFAULT_KEEP_ALIVE,
     FAL_SERVERLESS_DEFAULT_MAX_MULTIPLEXING,
+    FAL_SERVERLESS_DEFAULT_MIN_CONCURRENCY,
     Credentials,
     FalServerlessClient,
     FalServerlessConnection,
@@ -309,6 +310,7 @@ class FalServerlessHost(Host):
             "machine_type",
             "keep_alive",
             "max_concurrency",
+            "min_concurrency",
             "max_multiplexing",
             "setup_function",
             "metadata",
@@ -352,6 +354,7 @@ def register(
         )
         keep_alive = options.host.get("keep_alive", FAL_SERVERLESS_DEFAULT_KEEP_ALIVE)
         max_concurrency = options.host.get("max_concurrency")
+        min_concurrency = options.host.get("min_concurrency")
         max_multiplexing = options.host.get("max_multiplexing")
         base_image = options.host.get("_base_image", None)
         scheduler = options.host.get("_scheduler", None)
@@ -367,6 +370,7 @@ def register(
             scheduler_options=scheduler_options,
             max_multiplexing=max_multiplexing,
             max_concurrency=max_concurrency,
+            min_concurrency=min_concurrency,
         )
 
         partial_func = _prepare_partial_func(func)
@@ -416,6 +420,7 @@ def run(
         )
         keep_alive = options.host.get("keep_alive", FAL_SERVERLESS_DEFAULT_KEEP_ALIVE)
         max_concurrency = options.host.get("max_concurrency")
+        min_concurrency = options.host.get("min_concurrency")
         max_multiplexing = options.host.get("max_multiplexing")
         base_image = options.host.get("_base_image", None)
         scheduler = options.host.get("_scheduler", None)
@@ -432,6 +437,7 @@ def run(
             scheduler_options=scheduler_options,
             max_multiplexing=max_multiplexing,
             max_concurrency=max_concurrency,
+            min_concurrency=min_concurrency,
         )
 
         return_value = _UNSET
@@ -556,6 +562,7 @@ def function(
     machine_type: str = FAL_SERVERLESS_DEFAULT_MACHINE_TYPE,
     keep_alive: int = FAL_SERVERLESS_DEFAULT_KEEP_ALIVE,
     max_multiplexing: int = FAL_SERVERLESS_DEFAULT_MAX_MULTIPLEXING,
+    min_concurrency: int = FAL_SERVERLESS_DEFAULT_MIN_CONCURRENCY,
     setup_function: Callable[..., None] | None = None,
     _base_image: str | None = None,
     _scheduler: str | None = None,
@@ -581,6 +588,7 @@ def function(
     machine_type: str = FAL_SERVERLESS_DEFAULT_MACHINE_TYPE,
     keep_alive: int = FAL_SERVERLESS_DEFAULT_KEEP_ALIVE,
     max_multiplexing: int = FAL_SERVERLESS_DEFAULT_MAX_MULTIPLEXING,
+    min_concurrency: int = FAL_SERVERLESS_DEFAULT_MIN_CONCURRENCY,
     setup_function: Callable[..., None] | None = None,
     _base_image: str | None = None,
     _scheduler: str | None = None,
@@ -658,6 +666,7 @@ def function(
     machine_type: str = FAL_SERVERLESS_DEFAULT_MACHINE_TYPE,
     keep_alive: int = FAL_SERVERLESS_DEFAULT_KEEP_ALIVE,
     max_multiplexing: int = FAL_SERVERLESS_DEFAULT_MAX_MULTIPLEXING,
+    min_concurrency: int = FAL_SERVERLESS_DEFAULT_MIN_CONCURRENCY,
     setup_function: Callable[..., None] | None = None,
     _base_image: str | None = None,
     _scheduler: str | None = None,
@@ -688,6 +697,7 @@ def function(
     machine_type: str = FAL_SERVERLESS_DEFAULT_MACHINE_TYPE,
     keep_alive: int = FAL_SERVERLESS_DEFAULT_KEEP_ALIVE,
     max_multiplexing: int = FAL_SERVERLESS_DEFAULT_MAX_MULTIPLEXING,
+    min_concurrency: int = FAL_SERVERLESS_DEFAULT_MIN_CONCURRENCY,
     setup_function: Callable[..., None] | None = None,
     _base_image: str | None = None,
     _scheduler: str | None = None,

diff --git a/projects/fal/src/fal/cli.py b/projects/fal/src/fal/cli.py
@@ -365,6 +365,7 @@ def _alias_table(aliases: list[AliasInfo]):
     table.add_column("Alias")
     table.add_column("Revision")
     table.add_column("Auth")
+    table.add_column("Min Concurrency")
     table.add_column("Max Concurrency")
     table.add_column("Max Multiplexing")
     table.add_column("Keep Alive")
@@ -375,6 +376,7 @@ def _alias_table(aliases: list[AliasInfo]):
             app_alias.alias,
             app_alias.revision,
             app_alias.auth_mode,
+            str(app_alias.min_concurrency),
             str(app_alias.max_concurrency),
             str(app_alias.max_multiplexing),
             str(app_alias.keep_alive),
@@ -429,6 +431,7 @@ def alias_list(client: api.FalServerlessClient):
 @click.option("--keep-alive", "-k", type=int)
 @click.option("--max-multiplexing", "-m", type=int)
 @click.option("--max-concurrency", "-c", type=int)
+@click.option("--min-concurrency", type=int)
 # TODO: add auth_mode
 # @click.option(
 #     "--auth",
@@ -442,6 +445,7 @@ def alias_update(
     keep_alive: int | None,
     max_multiplexing: int | None,
     max_concurrency: int | None,
+    min_concurrency: int | None,
 ):
     with client.connect() as connection:
         if keep_alive is None and max_multiplexing is None and max_concurrency is None:
@@ -453,6 +457,7 @@ def alias_update(
             keep_alive=keep_alive,
             max_multiplexing=max_multiplexing,
             max_concurrency=max_concurrency,
+            min_concurrency=min_concurrency,
         )
         table = _alias_table([alias_info])
 

diff --git a/projects/fal/src/fal/sdk.py b/projects/fal/src/fal/sdk.py
@@ -27,6 +27,7 @@
 _DEFAULT_SERIALIZATION_METHOD = "dill"
 FAL_SERVERLESS_DEFAULT_KEEP_ALIVE = 10
 FAL_SERVERLESS_DEFAULT_MAX_MULTIPLEXING = 1
+FAL_SERVERLESS_DEFAULT_MIN_CONCURRENCY = 0
 
 log = get_logger(__name__)
 
@@ -188,6 +189,7 @@ class AliasInfo:
     max_concurrency: int
     max_multiplexing: int
     active_runners: int
+    min_concurrency: int
 
 
 @dataclass
@@ -272,6 +274,7 @@ def _from_grpc_alias_info(message: isolate_proto.AliasInfo) -> AliasInfo:
         max_concurrency=message.max_concurrency,
         max_multiplexing=message.max_multiplexing,
         active_runners=message.active_runners,
+        min_concurrency=message.min_concurrency,
     )
 
 
@@ -330,6 +333,7 @@ class MachineRequirements:
     scheduler_options: dict[str, Any] | None = None
     max_concurrency: int | None = None
     max_multiplexing: int | None = None
+    min_concurrency: int | None = None
 
 
 @dataclass
@@ -425,6 +429,7 @@ def register(
                     machine_requirements.scheduler_options or {}
                 ),
                 max_concurrency=machine_requirements.max_concurrency,
+                min_concurrency=machine_requirements.min_concurrency,
                 max_multiplexing=machine_requirements.max_multiplexing,
             )
         else:
@@ -462,12 +467,14 @@ def update_application(
         keep_alive: int | None = None,
         max_multiplexing: int | None = None,
         max_concurrency: int | None = None,
+        min_concurrency: int | None = None,
     ) -> AliasInfo:
         request = isolate_proto.UpdateApplicationRequest(
             application_name=application_name,
             keep_alive=keep_alive,
             max_multiplexing=max_multiplexing,
             max_concurrency=max_concurrency,
+            min_concurrency=min_concurrency,
         )
         res: isolate_proto.UpdateApplicationResult = self.stub.UpdateApplication(
             request
@@ -496,6 +503,7 @@ def run(
                 ),
                 max_concurrency=machine_requirements.max_concurrency,
                 max_multiplexing=machine_requirements.max_multiplexing,
+                min_concurrency=machine_requirements.min_concurrency,
             )
         else:
             wrapped_requirements = None

diff --git a/projects/isolate_proto/src/isolate_proto/controller.proto b/projects/isolate_proto/src/isolate_proto/controller.proto
@@ -148,6 +148,7 @@ message MachineRequirements {
     optional google.protobuf.Struct scheduler_options = 8;
     optional int32 max_multiplexing = 6;
     optional int32 max_concurrency = 9;
+    optional int32 min_concurrency = 10;
 }
 
 enum ApplicationAuthMode {
@@ -189,6 +190,7 @@ message UpdateApplicationRequest {
     optional int32 keep_alive = 2;
     optional int32 max_multiplexing = 3;
     optional int32 max_concurrency = 4;
+    optional int32 min_concurrency = 5;
 }
 
 message UpdateApplicationResult {
@@ -229,6 +231,7 @@ message AliasInfo {
     int32 max_multiplexing = 5;
     int32 keep_alive = 6;
     int32 active_runners = 7;
+    int32 min_concurrency = 8;
 }
 
 message SetSecretRequest {