gpu-mointor-vector-config

vector.toml

# Vector configuration for collecting NVIDIA GPU metrics

[api]
enabled = true
address = "0.0.0.0:8686"

# Source configuration for scraping Prometheus metrics from DCGM exporter
[sources.dcgm_metrics]
type = "prometheus_scrape"
endpoints = ["http://dcgm-exporter:9400/metrics"]
scrape_interval_secs = 15

# Transform configuration to process the metrics 
[transforms.gpu_metrics_processor]
type = "remap"
inputs = ["dcgm_metrics"]
source = '''
# Handle type checking properly
if exists(.metric_name) {
  # Convert to string with fallback
  .metric_name = to_string(.metric_name) ?? ""
  
  # Now we can safely check if it starts with DCGM
  if !starts_with(.metric_name, "DCGM") {
    # If not a DCGM metric, remove it by returning null
    null
  } else {
    # Add a timestamp if missing
    if !exists(.timestamp) {
      .timestamp = now()
    }
    # Keep the event by returning it
    .
  }
} else {
  # If no metric_name, add a type for debugging
  .metric_type = "unknown"
  .
}
'''

# Sink configuration to send metrics to Redpanda/Kafka
[sinks.kafka_output]
type = "kafka"
inputs = ["gpu_metrics_processor"]
bootstrap_servers = "redpanda:9092"
topic = "gpu-metrics"
encoding.codec = "json"