增加必要的数据记录

This commit is contained in:
xiuting.xu
2026-06-16 14:05:26 +08:00
parent fa19fc8722
commit b70bad4721
24 changed files with 1784 additions and 504 deletions

565
src/rtr/report.rs Normal file
View File

@@ -0,0 +1,565 @@
use std::fs;
use std::path::Path;
use std::sync::{Arc, RwLock};
use std::time::Instant;
use anyhow::{Context, Result};
use chrono::{DateTime, Utc};
use chrono_tz::Tz;
use serde::Serialize;
use tracing::warn;
use crate::rtr::cache::{CacheAvailability, SharedRtrCache, VersionReportStats};
use crate::rtr::config::format_timezone;
use crate::rtr::server::{RtrNotifier, RtrServiceStats, RtrTransportConnectionCounts};
use crate::source::pipeline::{DataQualityReport, SourceLoadReport};
#[derive(Clone, Serialize)]
pub struct ReportConfiguration {
source_refresh_interval_seconds: u64,
report_interval_seconds: u64,
max_delta: u8,
prune_delta_by_snapshot_size: bool,
strict_ccr_validation: bool,
timezone: String,
timing: TimingReport,
}
#[derive(Clone, Copy, Serialize)]
struct TimingReport {
refresh: u32,
retry: u32,
expire: u32,
}
impl ReportConfiguration {
pub fn new(
source_refresh_interval_seconds: u64,
report_interval_seconds: u64,
max_delta: u8,
prune_delta_by_snapshot_size: bool,
strict_ccr_validation: bool,
timezone: Tz,
timing: (u32, u32, u32),
) -> Self {
Self {
source_refresh_interval_seconds,
report_interval_seconds,
max_delta,
prune_delta_by_snapshot_size,
strict_ccr_validation,
timezone: format_timezone(timezone),
timing: TimingReport {
refresh: timing.0,
retry: timing.1,
expire: timing.2,
},
}
}
pub fn timezone(&self) -> Tz {
self.timezone
.parse::<Tz>()
.expect("serialized timezone should be a valid IANA timezone")
}
}
#[derive(Clone)]
pub struct ReportContext {
started_at: DateTime<Utc>,
started_instant: Instant,
timezone: Tz,
configuration: ReportConfiguration,
runtime: Arc<RwLock<RuntimeReportState>>,
}
#[derive(Debug, Clone, Default)]
struct RuntimeReportState {
source: Option<SourceLoadReport>,
data_quality: Option<DataQualityReport>,
refresh: RefreshReport,
}
#[derive(Debug, Clone, Serialize)]
struct RefreshReport {
last_attempt_at: Option<DateTime<Utc>>,
last_success_at: Option<DateTime<Utc>>,
last_changed_at: Option<DateTime<Utc>>,
status: &'static str,
changed: Option<bool>,
duration_ms: Option<u128>,
consecutive_failures: u64,
last_error: Option<String>,
}
#[derive(Serialize)]
struct RefreshReportView {
last_attempt_at: Option<DateTime<Tz>>,
last_success_at: Option<DateTime<Tz>>,
last_changed_at: Option<DateTime<Tz>>,
status: &'static str,
changed: Option<bool>,
duration_ms: Option<u128>,
consecutive_failures: u64,
last_error: Option<String>,
}
impl RefreshReportView {
fn from_report(report: RefreshReport, timezone: Tz) -> Self {
Self {
last_attempt_at: report
.last_attempt_at
.map(|time| to_report_time(time, timezone)),
last_success_at: report
.last_success_at
.map(|time| to_report_time(time, timezone)),
last_changed_at: report
.last_changed_at
.map(|time| to_report_time(time, timezone)),
status: report.status,
changed: report.changed,
duration_ms: report.duration_ms,
consecutive_failures: report.consecutive_failures,
last_error: report.last_error,
}
}
}
impl Default for RefreshReport {
fn default() -> Self {
Self {
last_attempt_at: None,
last_success_at: None,
last_changed_at: None,
status: "not_attempted",
changed: None,
duration_ms: None,
consecutive_failures: 0,
last_error: None,
}
}
}
impl ReportContext {
pub fn new(configuration: ReportConfiguration) -> Self {
let timezone = configuration.timezone();
Self {
started_at: Utc::now(),
started_instant: Instant::now(),
timezone,
configuration,
runtime: Arc::new(RwLock::new(RuntimeReportState::default())),
}
}
pub fn record_refresh_success(
&self,
attempted_at: DateTime<Utc>,
duration_ms: u128,
changed: bool,
source: SourceLoadReport,
data_quality: DataQualityReport,
) {
let mut state = self
.runtime
.write()
.unwrap_or_else(|poisoned| poisoned.into_inner());
state.source = Some(source);
state.data_quality = Some(data_quality);
state.refresh.last_attempt_at = Some(attempted_at);
state.refresh.last_success_at = Some(Utc::now());
if changed {
state.refresh.last_changed_at = Some(Utc::now());
}
state.refresh.status = "success";
state.refresh.changed = Some(changed);
state.refresh.duration_ms = Some(duration_ms);
state.refresh.consecutive_failures = 0;
state.refresh.last_error = None;
}
pub fn record_refresh_unchanged(&self, attempted_at: DateTime<Utc>, duration_ms: u128) {
let mut state = self
.runtime
.write()
.unwrap_or_else(|poisoned| poisoned.into_inner());
state.refresh.last_attempt_at = Some(attempted_at);
state.refresh.last_success_at = Some(Utc::now());
state.refresh.status = "success";
state.refresh.changed = Some(false);
state.refresh.duration_ms = Some(duration_ms);
state.refresh.consecutive_failures = 0;
state.refresh.last_error = None;
}
pub fn record_refresh_failure(
&self,
attempted_at: DateTime<Utc>,
duration_ms: u128,
error: &anyhow::Error,
) {
let mut state = self
.runtime
.write()
.unwrap_or_else(|poisoned| poisoned.into_inner());
state.refresh.last_attempt_at = Some(attempted_at);
state.refresh.status = "failed";
state.refresh.changed = None;
state.refresh.duration_ms = Some(duration_ms);
state.refresh.consecutive_failures = state.refresh.consecutive_failures.saturating_add(1);
state.refresh.last_error = Some(error.to_string());
}
pub fn write_or_warn(
&self,
report_dir: &Path,
phase: &str,
shared_cache: &SharedRtrCache,
notifier: &RtrNotifier,
service_stats: &RtrServiceStats,
) {
if let Err(err) = self.write(report_dir, phase, shared_cache, notifier, service_stats) {
warn!(
"failed to write RTR report to {}: {:?}",
report_dir.display(),
err
);
}
}
fn write(
&self,
report_dir: &Path,
phase: &str,
shared_cache: &SharedRtrCache,
notifier: &RtrNotifier,
service_stats: &RtrServiceStats,
) -> Result<()> {
let cache = shared_cache.load_full();
let runtime = self
.runtime
.read()
.unwrap_or_else(|poisoned| poisoned.into_inner())
.clone();
let availability = match cache.availability() {
CacheAvailability::Ready => "ready",
CacheAvailability::NoDataAvailable => "no_data_available",
};
let active_connections = service_stats.active_connections();
let connections_by_transport = service_stats.transport_connections();
let max_connections = service_stats.max_connections();
let report = RtrReport {
schema_version: 2,
generated_at: self.report_now(),
phase: phase.to_string(),
service: ServiceReport {
started_at: self.to_report_time(self.started_at),
uptime_seconds: self.started_instant.elapsed().as_secs(),
active_connections,
connections_by_transport: TransportConnectionReport::from(connections_by_transport),
session_listeners: notifier.listener_count(),
max_connections,
connection_utilization: active_connections as f64 / max_connections as f64,
},
process: ProcessReport {
rss_mib: current_rss_mib(),
},
source: runtime
.source
.map(|source| SourceLoadReportView::from_report(source, self.timezone)),
refresh: RefreshReportView::from_report(runtime.refresh, self.timezone),
data_quality: runtime.data_quality,
configuration: self.configuration.clone(),
cache: CacheReport {
availability,
created_at: self.to_report_time(cache.created_at().utc()),
last_update_begin: self.to_report_time(cache.last_update_begin().utc()),
last_update_end: self.to_report_time(cache.last_update_end().utc()),
versions: cache.version_report_stats(),
},
};
fs::create_dir_all(report_dir)
.with_context(|| format!("create report directory {}", report_dir.display()))?;
let target = report_dir.join("rtr-server.json");
let temporary = report_dir.join(".rtr-server.json.tmp");
let json = serde_json::to_vec_pretty(&report).context("serialize RTR report")?;
fs::write(&temporary, json)
.with_context(|| format!("write temporary report {}", temporary.display()))?;
if let Err(err) = fs::rename(&temporary, &target) {
if target.exists() {
fs::remove_file(&target)
.with_context(|| format!("replace existing report {}", target.display()))?;
fs::rename(&temporary, &target)
.with_context(|| format!("move report into {}", target.display()))?;
} else {
return Err(err).with_context(|| {
format!(
"move temporary report {} into {}",
temporary.display(),
target.display()
)
});
}
}
Ok(())
}
fn report_now(&self) -> DateTime<Tz> {
self.to_report_time(Utc::now())
}
fn to_report_time(&self, time: DateTime<Utc>) -> DateTime<Tz> {
time.with_timezone(&self.timezone)
}
}
#[derive(Serialize)]
struct RtrReport {
schema_version: u16,
generated_at: DateTime<Tz>,
phase: String,
service: ServiceReport,
process: ProcessReport,
source: Option<SourceLoadReportView>,
refresh: RefreshReportView,
data_quality: Option<DataQualityReport>,
configuration: ReportConfiguration,
cache: CacheReport,
}
#[derive(Serialize)]
struct ServiceReport {
started_at: DateTime<Tz>,
uptime_seconds: u64,
active_connections: usize,
connections_by_transport: TransportConnectionReport,
session_listeners: usize,
max_connections: usize,
connection_utilization: f64,
}
#[derive(Serialize)]
struct SourceLoadReportView {
ccr_file: String,
ccr_file_size_bytes: u64,
ccr_modified_at: Option<DateTime<Tz>>,
ccr_produced_at: Option<String>,
slurm_enabled: bool,
slurm_file_count: usize,
slurm_files: Vec<String>,
slurm_version: Option<u32>,
}
impl SourceLoadReportView {
fn from_report(report: SourceLoadReport, timezone: Tz) -> Self {
Self {
ccr_file: report.ccr_file,
ccr_file_size_bytes: report.ccr_file_size_bytes,
ccr_modified_at: report
.ccr_modified_at
.map(|time| to_report_time(time, timezone)),
ccr_produced_at: report.ccr_produced_at,
slurm_enabled: report.slurm_enabled,
slurm_file_count: report.slurm_file_count,
slurm_files: report.slurm_files,
slurm_version: report.slurm_version,
}
}
}
#[derive(Serialize)]
struct TransportConnectionReport {
tcp: usize,
tls: usize,
ssh: usize,
}
impl From<RtrTransportConnectionCounts> for TransportConnectionReport {
fn from(counts: RtrTransportConnectionCounts) -> Self {
Self {
tcp: counts.tcp,
tls: counts.tls,
ssh: counts.ssh,
}
}
}
#[derive(Serialize)]
struct ProcessReport {
rss_mib: Option<u64>,
}
#[derive(Serialize)]
struct CacheReport {
availability: &'static str,
created_at: DateTime<Tz>,
last_update_begin: DateTime<Tz>,
last_update_end: DateTime<Tz>,
versions: [VersionReportStats; 3],
}
pub fn to_report_time(time: DateTime<Utc>, timezone: Tz) -> DateTime<Tz> {
time.with_timezone(&timezone)
}
pub fn current_rss_mib() -> Option<u64> {
let status = fs::read_to_string("/proc/self/status").ok()?;
let vmrss_line = status.lines().find(|line| line.starts_with("VmRSS:"))?;
let kb = vmrss_line
.split_whitespace()
.nth(1)
.and_then(|value| value.parse::<u64>().ok())?;
Some(kb / 1024)
}
#[cfg(test)]
mod tests {
use std::sync::Arc;
use crate::rtr::cache::RtrCache;
use crate::rtr::server::RtrService;
use crate::source::pipeline::{
DataQualityReport, PayloadTypeCounts, SlurmRuleCounts, SourceLoadReport,
};
use arc_swap::ArcSwap;
use serde_json::Value;
use super::*;
fn test_configuration() -> ReportConfiguration {
ReportConfiguration::new(
300,
60,
100,
false,
false,
chrono_tz::Asia::Shanghai,
(3600, 600, 7200),
)
}
#[test]
fn write_report_creates_parseable_json() {
let temp = tempfile::tempdir().unwrap();
let report_dir = temp.path().join("report");
let shared_cache = Arc::new(ArcSwap::from_pointee(RtrCache::default()));
let service = RtrService::new(shared_cache.clone());
let notifier = service.notifier();
let context = ReportContext::new(test_configuration());
context
.write(
&report_dir,
"test",
&shared_cache,
&notifier,
&service.stats(),
)
.unwrap();
let report: Value =
serde_json::from_slice(&fs::read(report_dir.join("rtr-server.json")).unwrap()).unwrap();
assert_eq!(report["schema_version"], 2);
assert_eq!(report["phase"], "test");
assert_report_time_offset(&report["generated_at"]);
assert_report_time_offset(&report["service"]["started_at"]);
assert_report_time_offset(&report["cache"]["created_at"]);
assert_eq!(report["cache"]["availability"], "ready");
assert_eq!(report["refresh"]["status"], "not_attempted");
assert!(report["source"].is_null());
assert!(report["data_quality"].is_null());
assert_eq!(
report["configuration"]["source_refresh_interval_seconds"],
300
);
assert_eq!(report["service"]["max_connections"], 1024);
assert_eq!(report["service"]["active_connections"], 0);
assert_eq!(report["service"]["connections_by_transport"]["tcp"], 0);
assert_eq!(report["service"]["connections_by_transport"]["tls"], 0);
assert_eq!(report["service"]["connections_by_transport"]["ssh"], 0);
assert_eq!(report["cache"]["versions"].as_array().unwrap().len(), 3);
assert_eq!(report["cache"]["versions"][2]["snapshot"]["total"], 0);
assert!(!report_dir.join(".rtr-server.json.tmp").exists());
}
#[test]
fn refresh_failure_preserves_last_successful_source_data() {
let temp = tempfile::tempdir().unwrap();
let report_dir = temp.path().join("report");
let shared_cache = Arc::new(ArcSwap::from_pointee(RtrCache::default()));
let service = RtrService::new(shared_cache.clone());
let notifier = service.notifier();
let context = ReportContext::new(test_configuration());
let source = SourceLoadReport {
ccr_file: "data/example.ccr".to_string(),
ccr_file_size_bytes: 123,
ccr_modified_at: Some(Utc::now()),
ccr_produced_at: Some("20260615000000Z".to_string()),
slurm_enabled: true,
slurm_file_count: 1,
slurm_files: vec!["policy.slurm".to_string()],
slurm_version: Some(2),
};
let quality = DataQualityReport {
ccr_input: PayloadTypeCounts {
total: 11,
vrp: 10,
router_key: 0,
aspa: 1,
},
invalid: PayloadTypeCounts::default(),
before_slurm: PayloadTypeCounts {
total: 11,
vrp: 10,
router_key: 0,
aspa: 1,
},
after_slurm: PayloadTypeCounts {
total: 10,
vrp: 9,
router_key: 0,
aspa: 1,
},
slurm_filters: SlurmRuleCounts {
prefix: 1,
router_key: 0,
aspa: 0,
},
slurm_assertions: SlurmRuleCounts::default(),
};
context.record_refresh_success(Utc::now(), 12, true, source, quality);
context.record_refresh_failure(Utc::now(), 5, &anyhow::anyhow!("source unavailable"));
context
.write(
&report_dir,
"refresh_failed",
&shared_cache,
&notifier,
&service.stats(),
)
.unwrap();
let report: Value =
serde_json::from_slice(&fs::read(report_dir.join("rtr-server.json")).unwrap()).unwrap();
assert_eq!(report["source"]["ccr_file"], "data/example.ccr");
assert_report_time_offset(&report["source"]["ccr_modified_at"]);
assert_eq!(report["data_quality"]["after_slurm"]["total"], 10);
assert_eq!(report["refresh"]["status"], "failed");
assert_eq!(report["refresh"]["consecutive_failures"], 1);
assert_eq!(report["refresh"]["last_error"], "source unavailable");
assert!(!report["refresh"]["last_success_at"].is_null());
assert_report_time_offset(&report["refresh"]["last_success_at"]);
}
fn assert_report_time_offset(value: &Value) {
let value = value.as_str().expect("report time should be a string");
assert!(
value.ends_with("+08:00"),
"report time should use +08:00 offset, got {value}"
);
}
}