wehub-resource-sync
58 lines
2.6 KiB
SQL
58 lines
2.6 KiB
SQL
-- +goose Up
|
|
|
|
-- Pre-aggregated model performance metrics with no tenant information.
|
|
-- Used for cross-tenant model comparisons in the Model Registry.
|
|
-- Aggregated per minute for high-resolution model performance tracking.
|
|
CREATE TABLE IF NOT EXISTS trigger_dev.llm_model_aggregates_v1
|
|
(
|
|
response_model String,
|
|
base_response_model String DEFAULT '',
|
|
gen_ai_system LowCardinality(String),
|
|
minute DateTime,
|
|
|
|
-- Counts & totals (SimpleAggregateFunction for sum)
|
|
call_count SimpleAggregateFunction(sum, UInt64),
|
|
total_input_tokens SimpleAggregateFunction(sum, UInt64),
|
|
total_output_tokens SimpleAggregateFunction(sum, UInt64),
|
|
total_cost SimpleAggregateFunction(sum, Float64),
|
|
|
|
-- Performance quantiles (AggregateFunction for merge across parts)
|
|
ttfc_quantiles AggregateFunction(quantiles(0.5, 0.9, 0.95, 0.99), Float64),
|
|
tps_quantiles AggregateFunction(quantiles(0.5, 0.9, 0.95, 0.99), Float64),
|
|
duration_quantiles AggregateFunction(quantiles(0.5, 0.9, 0.95, 0.99), UInt64),
|
|
|
|
-- Finish reason distribution
|
|
finish_reason_counts SimpleAggregateFunction(sumMap, Map(String, UInt64))
|
|
)
|
|
ENGINE = AggregatingMergeTree
|
|
PARTITION BY toYYYYMM(minute)
|
|
ORDER BY (response_model, base_response_model, gen_ai_system, minute)
|
|
TTL toDate(minute) + INTERVAL 365 DAY
|
|
SETTINGS ttl_only_drop_parts = 1;
|
|
|
|
-- Materialized view that feeds the aggregate table from llm_metrics_v1.
|
|
-- Strips all tenant-specific columns (org, project, env, run, span, trace).
|
|
-- base_response_model comes from the source table (populated during event enrichment).
|
|
CREATE MATERIALIZED VIEW IF NOT EXISTS trigger_dev.llm_model_aggregates_mv_v1
|
|
TO trigger_dev.llm_model_aggregates_v1
|
|
AS SELECT
|
|
response_model,
|
|
base_response_model,
|
|
gen_ai_system,
|
|
toStartOfMinute(start_time) AS minute,
|
|
count() AS call_count,
|
|
sum(input_tokens) AS total_input_tokens,
|
|
sum(output_tokens) AS total_output_tokens,
|
|
sum(total_cost) AS total_cost,
|
|
quantilesStateIf(0.5, 0.9, 0.95, 0.99)(ms_to_first_chunk, ms_to_first_chunk > 0) AS ttfc_quantiles,
|
|
quantilesStateIf(0.5, 0.9, 0.95, 0.99)(tokens_per_second, tokens_per_second > 0) AS tps_quantiles,
|
|
quantilesState(0.5, 0.9, 0.95, 0.99)(duration) AS duration_quantiles,
|
|
sumMap(map(finish_reason, toUInt64(1))) AS finish_reason_counts
|
|
FROM trigger_dev.llm_metrics_v1
|
|
WHERE response_model != ''
|
|
GROUP BY response_model, base_response_model, gen_ai_system, minute;
|
|
|
|
-- +goose Down
|
|
DROP TABLE IF EXISTS trigger_dev.llm_model_aggregates_mv_v1;
|
|
DROP TABLE IF EXISTS trigger_dev.llm_model_aggregates_v1;
|