Parse Postgres and pgAudit logs using the OTel Collector

Postgres can log in two structured formats: CSV and JSON since Postgres 15.
The two formats are very similar semantically, so this parses them in a
shared OTTL transform processor.

Co-authored-by: Chris Bandy <[email protected]>
Issue: PGO-2033
Issue: PGO-2065

internal/collector/instance.go

@@ -40,7 +40,7 @@ func AddToPod(
 	volumeMounts []corev1.VolumeMount,
 	sqlQueryPassword string,
 ) {
-	if !feature.Enabled(ctx, feature.OpenTelemetryMetrics) {
+	if !(feature.Enabled(ctx, feature.OpenTelemetryLogs) || feature.Enabled(ctx, feature.OpenTelemetryMetrics)) {
 		return
 	}
 
@@ -67,10 +67,22 @@ func AddToPod(
 	container := corev1.Container{
 		Name: naming.ContainerCollector,
 
-		Image:           "ghcr.io/open-telemetry/opentelemetry-collector-releases/opentelemetry-collector-contrib:0.116.1",
+		Image:           "ghcr.io/open-telemetry/opentelemetry-collector-releases/opentelemetry-collector-contrib:0.117.0",
 		ImagePullPolicy: inCluster.Spec.ImagePullPolicy,
 		Command:         []string{"/otelcol-contrib", "--config", "/etc/otel-collector/config.yaml"},
 		Env: []corev1.EnvVar{
+			{
+				Name: "K8S_POD_NAMESPACE",
+				ValueFrom: &corev1.EnvVarSource{FieldRef: &corev1.ObjectFieldSelector{
+					FieldPath: "metadata.namespace",
+				}},
+			},
+			{
+				Name: "K8S_POD_NAME",
+				ValueFrom: &corev1.EnvVarSource{FieldRef: &corev1.ObjectFieldSelector{
+					FieldPath: "metadata.name",
+				}},
+			},
 			{
 				Name:  "PGPASSWORD",
 				Value: sqlQueryPassword,
@@ -81,11 +93,13 @@ func AddToPod(
 		VolumeMounts:    append(volumeMounts, configVolumeMount),
 	}
 
-	container.Ports = []corev1.ContainerPort{{
-		ContainerPort: int32(8889),
-		Name:          "otel-metrics",
-		Protocol:      corev1.ProtocolTCP,
-	}}
+	if feature.Enabled(ctx, feature.OpenTelemetryMetrics) {
+		container.Ports = []corev1.ContainerPort{{
+			ContainerPort: int32(8889),
+			Name:          "otel-metrics",
+			Protocol:      corev1.ProtocolTCP,
+		}}
+	}
 
 	outPod.Containers = append(outPod.Containers, container)
 	outPod.Volumes = append(outPod.Volumes, configVolume)

internal/collector/postgres.go

@@ -6,14 +6,200 @@ package collector
 
 import (
 	"context"
+	_ "embed"
+	"encoding/json"
+	"fmt"
+	"slices"
 
+	"github.com/crunchydata/postgres-operator/internal/feature"
+	"github.com/crunchydata/postgres-operator/internal/naming"
+	"github.com/crunchydata/postgres-operator/internal/postgres"
 	"github.com/crunchydata/postgres-operator/pkg/apis/postgres-operator.crunchydata.com/v1beta1"
 )
 
-func NewConfigForPostgresPod(ctx context.Context, inCluster *v1beta1.PostgresCluster) *Config {
+func NewConfigForPostgresPod(ctx context.Context,
+	inCluster *v1beta1.PostgresCluster,
+	outParameters *postgres.Parameters,
+) *Config {
 	config := NewConfig()
 
 	EnablePatroniMetrics(ctx, inCluster, config)
+	EnablePostgresLogging(ctx, inCluster, config, outParameters)
 
 	return config
 }
+
+// The contents of "postgres_logs_transforms.yaml" as JSON.
+// See: https://pkg.go.dev/embed
+//
+//go:embed "generated/postgres_logs_transforms.json"
+var postgresLogsTransforms json.RawMessage
+
+// postgresCSVNames returns the names of fields in the CSV logs for version.
+func postgresCSVNames(version int) string {
+	// JSON is the preferred format, so use those names.
+	// https://www.postgresql.org/docs/current/runtime-config-logging.html#RUNTIME-CONFIG-LOGGING-JSONLOG
+
+	// https://www.postgresql.org/docs/8.3/runtime-config-logging.html#RUNTIME-CONFIG-LOGGING-CSVLOG
+	names := `timestamp,user,dbname,pid` +
+		`,connection_from` + // NOTE: this contains the JSON "remote_host" and "remote_port" values
+		`,session_id,line_num,ps,session_start,vxid,txid` +
+		`,error_severity,state_code,message,detail,hint` +
+		`,internal_query,internal_position,context,statement,cursor_position` +
+		`,location` // NOTE: this contains the JSON "func_name", "file_name", and "file_line_num" values
+
+	// https://www.postgresql.org/docs/9.0/runtime-config-logging.html#RUNTIME-CONFIG-LOGGING-CSVLOG
+	if version >= 9 {
+		names += `,application_name`
+	}
+
+	// https://www.postgresql.org/docs/13/runtime-config-logging.html#RUNTIME-CONFIG-LOGGING-CSVLOG
+	if version >= 13 {
+		names += `,backend_type`
+	}
+
+	// https://www.postgresql.org/docs/14/runtime-config-logging.html#RUNTIME-CONFIG-LOGGING-CSVLOG
+	if version >= 14 {
+		names += `,leader_pid,query_id`
+	}
+
+	return names
+}
+
+func EnablePostgresLogging(
+	ctx context.Context,
+	inCluster *v1beta1.PostgresCluster,
+	outConfig *Config,
+	outParameters *postgres.Parameters,
+) {
+	if feature.Enabled(ctx, feature.OpenTelemetryLogs) {
+		directory := postgres.LogDirectory()
+
+		// https://www.postgresql.org/docs/current/runtime-config-logging.html
+		outParameters.Mandatory.Add("logging_collector", "on")
+		outParameters.Mandatory.Add("log_directory", directory)
+
+		// PostgreSQL v8.3 adds support for CSV logging, and
+		// PostgreSQL v15 adds support for JSON logging. The latter is preferred
+		// because newlines are escaped as "\n", U+005C + U+006E.
+		if inCluster.Spec.PostgresVersion < 15 {
+			outParameters.Mandatory.Add("log_destination", "csvlog")
+		} else {
+			outParameters.Mandatory.Add("log_destination", "jsonlog")
+		}
+
+		// Keep seven days of logs named for the day of the week;
+		// this has been the default produced by `initdb` for some time now.
+		// NOTE: The automated portions of log_filename are *entirely* based
+		// on time. There is no spelling that is guaranteed to be unique or
+		// monotonically increasing.
+		//
+		// TODO(logs): Limit the size/bytes of logs without losing messages;
+		// probably requires another process that deletes the oldest files.
+		//
+		// The ".log" suffix is replaced by ".json" for JSON log files.
+		outParameters.Mandatory.Add("log_filename", "postgresql-%a.log")
+		outParameters.Mandatory.Add("log_file_mode", "0660")
+		outParameters.Mandatory.Add("log_rotation_age", "1d")
+		outParameters.Mandatory.Add("log_rotation_size", "0")
+		outParameters.Mandatory.Add("log_truncate_on_rotation", "on")
+
+		// Log in a timezone that the OpenTelemetry Collector will understand.
+		outParameters.Mandatory.Add("log_timezone", "UTC")
+
+		// Keep track of what log records and files have been processed.
+		// Use a subdirectory of the logs directory to stay within the same failure domain.
+		//
+		// https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/-/extension/storage/filestorage#readme
+		outConfig.Extensions["file_storage/postgres_logs"] = map[string]any{
+			"directory":        directory + "/receiver",
+			"create_directory": true,
+			"fsync":            true,
+		}
+
+		// TODO(postgres-14): We can stop parsing CSV logs when 14 is EOL.
+		// https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/-/receiver/filelogreceiver#readme
+		outConfig.Receivers["filelog/postgres_csvlog"] = map[string]any{
+			// Read the CSV files and keep track of what has been processed.
+			"include": []string{directory + "/*.csv"},
+			"storage": "file_storage/postgres_logs",
+
+			// Postgres does not escape newlines in its CSV log format. Search for
+			// the beginning of every record, starting with an unquoted timestamp.
+			// The 2nd through 5th fields are optional, so match through to the 7th field.
+			// This should do a decent job of not matching the middle of some SQL statement.
+			//
+			// The number of fields has changed over the years, but the first few
+			// are always formatted the same way.
+			//
+			// NOTE: This regexp is invoked in multi-line mode. https://go.dev/s/re2syntax
+			"multiline": map[string]string{
+				"line_start_pattern": `^\d{4}-\d\d-\d\d \d\d:\d\d:\d\d.\d{3} UTC` + // 1st: timestamp
+					`,(?:"[_\D](?:[^"]|"")*")?` + //  2nd: user name
+					`,(?:"[_\D](?:[^"]|"")*")?` + //  3rd: database name
+					`,\d*,(?:"(?:[^"]|"")+")?` + //   4–5th: process id, connection
+					`,[0-9a-f]+[.][0-9a-f]+,\d+,`, // 6–7th: session id, session line
+			},
+
+			// Differentiate these from the JSON ones below.
+			"operators": []map[string]any{
+				{"type": "move", "from": "body", "to": "body.original"},
+				{"type": "add", "field": "body.format", "value": "csv"},
+				{"type": "add", "field": "body.headers", "value": postgresCSVNames(inCluster.Spec.PostgresVersion)},
+			},
+		}
+
+		// https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/-/receiver/filelogreceiver#readme
+		outConfig.Receivers["filelog/postgres_jsonlog"] = map[string]any{
+			// Read the JSON files and keep track of what has been processed.
+			"include": []string{directory + "/*.json"},
+			"storage": "file_storage/postgres_logs",
+
+			// Differentiate these from the CSV ones above.
+			// TODO(postgres-14): We can stop parsing CSV logs when 14 is EOL.
+			"operators": []map[string]any{
+				{"type": "move", "from": "body", "to": "body.original"},
+				{"type": "add", "field": "body.format", "value": "json"},
+			},
+		}
+
+		// https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/-/processor/resourceprocessor#readme
+		outConfig.Processors["resource/postgres"] = map[string]any{
+			"attributes": []map[string]any{
+				// Container and Namespace names need no escaping because they are DNS labels.
+				// Pod names need no escaping because they are DNS subdomains.
+				//
+				// https://kubernetes.io/docs/concepts/overview/working-with-objects/names
+				// https://github.com/open-telemetry/semantic-conventions/blob/v1.29.0/docs/resource/k8s.md
+				{"action": "insert", "key": "k8s.container.name", "value": naming.ContainerDatabase},
+				{"action": "insert", "key": "k8s.namespace.name", "value": "${env:K8S_POD_NAMESPACE}"},
+				{"action": "insert", "key": "k8s.pod.name", "value": "${env:K8S_POD_NAME}"},
+
+				// https://github.com/open-telemetry/semantic-conventions/blob/v1.29.0/docs/database#readme
+				{"action": "insert", "key": "db.system", "value": "postgresql"},
+				{"action": "insert", "key": "db.version", "value": fmt.Sprint(inCluster.Spec.PostgresVersion)},
+			},
+		}
+
+		// https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/-/processor/transformprocessor#readme
+		outConfig.Processors["transform/postgres_logs"] = map[string]any{
+			"log_statements": slices.Clone(postgresLogsTransforms),
+		}
+
+		outConfig.Pipelines["logs/postgres"] = Pipeline{
+			Extensions: []ComponentID{"file_storage/postgres_logs"},
+			// TODO(logs): Choose only one receiver, maybe?
+			Receivers: []ComponentID{
+				"filelog/postgres_csvlog",
+				"filelog/postgres_jsonlog",
+			},
+			Processors: []ComponentID{
+				"resource/postgres",
+				"transform/postgres_logs",
+				SubSecondBatchProcessor,
+				CompactingProcessor,
+			},
+			Exporters: []ComponentID{DebugExporter},
+		}
+	}
+}