apunkt/plano
1
0
Fork 0
mirror of https://github.com/katanemo/plano.git synced 2026-06-29 15:49:40 +02:00
Code Issues Projects Releases Packages Wiki Activity

update observability and dashboards

Browse source 
- spin up individual clusters for every endpoint so envoy can start tracking its latency
- update dashboards to show individual clusters
This commit is contained in:
Adil Hafeez 2024-10-18 12:35:57 -07:00
parent 6cd05572c4
commit 3796ee8773
5 changed files with 450 additions and 562 deletions
Download patch file Download diff file Expand all files Collapse all files

6
crates/common/src/consts.rs
View file

@ -6,9 +6,13 @@ pub const RATELIMIT_SELECTOR_HEADER_KEY: &str = "x-arch-ratelimit-selector";
pub const SYSTEM_ROLE: &str = "system";
pub const USER_ROLE: &str = "user";
pub const GPT_35_TURBO: &str = "gpt-3.5-turbo";
pub const ARC_FC_CLUSTER: &str = "arch_fc";
pub const ARCH_FC_REQUEST_TIMEOUT_MS: u64 = 120000; // 2 minutes
pub const MODEL_SERVER_NAME: &str = "model_server";
pub const ZEROSHOT_INTERNAL_HOST: &str = "zeroshot";
pub const ARCH_FC_INTERNAL_HOST: &str = "arch_fc";
pub const HALLUCINATION_INTERNAL_HOST: &str = "hallucination";
pub const EMBEDDINGS_INTERNAL_HOST: &str = "embeddings";
pub const GUARD_INTERNAL_HOST: &str = "guard";
pub const ARCH_ROUTING_HEADER: &str = "x-arch-llm-provider";
pub const ARCH_MESSAGES_KEY: &str = "arch_messages";
pub const ARCH_PROVIDER_HINT_HEADER: &str = "x-arch-llm-provider-hint";

7
crates/prompt_gateway/src/prompt_filter_context.rs
View file

@ -1,10 +1,9 @@
use crate::prompt_stream_context::PromptStreamContext;
use common::common_types::EmbeddingType;
use common::configuration::{Configuration, GatewayMode, Overrides, PromptGuards, PromptTarget};
use common::consts::ARCH_INTERNAL_CLUSTER_NAME;
use common::consts::{ARCH_INTERNAL_CLUSTER_NAME, EMBEDDINGS_INTERNAL_HOST};
use common::consts::ARCH_UPSTREAM_HOST_HEADER;
use common::consts::DEFAULT_EMBEDDING_MODEL;
use common::consts::MODEL_SERVER_NAME;
use common::embeddings::{
CreateEmbeddingRequest, CreateEmbeddingRequestInput, CreateEmbeddingResponse,
};
@ -105,10 +104,10 @@ impl PromptGatewayFilterContext {
ARCH_INTERNAL_CLUSTER_NAME,
"/embeddings",
vec![
(ARCH_UPSTREAM_HOST_HEADER, MODEL_SERVER_NAME),
(ARCH_UPSTREAM_HOST_HEADER, EMBEDDINGS_INTERNAL_HOST),
(":method", "POST"),
(":path", "/embeddings"),
(":authority", MODEL_SERVER_NAME),
(":authority", EMBEDDINGS_INTERNAL_HOST),
("content-type", "application/json"),
("x-envoy-upstream-rq-timeout-ms", "60000"),
],

26
crates/prompt_gateway/src/prompt_stream_context.rs
View file

@ -12,11 +12,7 @@ use common::common_types::{
};
use common::configuration::{Overrides, PromptGuards, PromptTarget};
use common::consts::{
ARCH_FC_MODEL_NAME, ARCH_FC_REQUEST_TIMEOUT_MS, ARCH_INTERNAL_CLUSTER_NAME, ARCH_MESSAGES_KEY,
ARCH_MODEL_PREFIX, ARCH_STATE_HEADER, ARCH_UPSTREAM_HOST_HEADER, ARC_FC_CLUSTER,
CHAT_COMPLETIONS_PATH, DEFAULT_EMBEDDING_MODEL, DEFAULT_HALLUCINATED_THRESHOLD,
DEFAULT_INTENT_MODEL, DEFAULT_PROMPT_TARGET_THRESHOLD, GPT_35_TURBO, MODEL_SERVER_NAME,
REQUEST_ID_HEADER, SYSTEM_ROLE, USER_ROLE,
ARCH_FC_MODEL_NAME, ARCH_FC_REQUEST_TIMEOUT_MS, ARCH_INTERNAL_CLUSTER_NAME, ARCH_MESSAGES_KEY, ARCH_MODEL_PREFIX, ARCH_STATE_HEADER, ARCH_UPSTREAM_HOST_HEADER, ARCH_FC_INTERNAL_HOST, CHAT_COMPLETIONS_PATH, DEFAULT_EMBEDDING_MODEL, DEFAULT_HALLUCINATED_THRESHOLD, DEFAULT_INTENT_MODEL, DEFAULT_PROMPT_TARGET_THRESHOLD, EMBEDDINGS_INTERNAL_HOST, GPT_35_TURBO, GUARD_INTERNAL_HOST, HALLUCINATION_INTERNAL_HOST, REQUEST_ID_HEADER, SYSTEM_ROLE, USER_ROLE, ZEROSHOT_INTERNAL_HOST
};
use common::embeddings::{
CreateEmbeddingRequest, CreateEmbeddingRequestInput, CreateEmbeddingResponse,
@ -239,10 +235,10 @@ impl PromptStreamContext {
};
let mut headers = vec![
(ARCH_UPSTREAM_HOST_HEADER, MODEL_SERVER_NAME),
(ARCH_UPSTREAM_HOST_HEADER, ZEROSHOT_INTERNAL_HOST),
(":method", "POST"),
(":path", "/zeroshot"),
(":authority", MODEL_SERVER_NAME),
(":authority", ZEROSHOT_INTERNAL_HOST),
("content-type", "application/json"),
("x-envoy-max-retries", "3"),
("x-envoy-upstream-rq-timeout-ms", "60000"),
@ -545,9 +541,9 @@ impl PromptStreamContext {
let mut headers = vec![
(":method", "POST"),
(ARCH_UPSTREAM_HOST_HEADER, ARC_FC_CLUSTER),
(ARCH_UPSTREAM_HOST_HEADER, ARCH_FC_INTERNAL_HOST),
(":path", "/v1/chat/completions"),
(":authority", ARC_FC_CLUSTER),
(":authority", ARCH_FC_INTERNAL_HOST),
("content-type", "application/json"),
("x-envoy-max-retries", "3"),
("x-envoy-upstream-rq-timeout-ms", timeout_str.as_str()),
@ -698,10 +694,10 @@ impl PromptStreamContext {
};
let mut headers = vec![
(ARCH_UPSTREAM_HOST_HEADER, MODEL_SERVER_NAME),
(ARCH_UPSTREAM_HOST_HEADER, HALLUCINATION_INTERNAL_HOST),
(":method", "POST"),
(":path", "/hallucination"),
(":authority", MODEL_SERVER_NAME),
(":authority", HALLUCINATION_INTERNAL_HOST),
("content-type", "application/json"),
("x-envoy-max-retries", "3"),
("x-envoy-upstream-rq-timeout-ms", "60000"),
@ -921,10 +917,10 @@ impl PromptStreamContext {
};
let mut headers = vec![
(ARCH_UPSTREAM_HOST_HEADER, MODEL_SERVER_NAME),
(ARCH_UPSTREAM_HOST_HEADER, EMBEDDINGS_INTERNAL_HOST),
(":method", "POST"),
(":path", "/embeddings"),
(":authority", MODEL_SERVER_NAME),
(":authority", EMBEDDINGS_INTERNAL_HOST),
("content-type", "application/json"),
("x-envoy-max-retries", "3"),
("x-envoy-upstream-rq-timeout-ms", "60000"),
@ -1177,10 +1173,10 @@ impl HttpContext for PromptStreamContext {
};
let mut headers = vec![
(ARCH_UPSTREAM_HOST_HEADER, MODEL_SERVER_NAME),
(ARCH_UPSTREAM_HOST_HEADER, GUARD_INTERNAL_HOST),
(":method", "POST"),
(":path", "/guard"),
(":authority", MODEL_SERVER_NAME),
(":authority", GUARD_INTERNAL_HOST),
("content-type", "application/json"),
("x-envoy-max-retries", "3"),
("x-envoy-upstream-rq-timeout-ms", "60000"),