Serve the /api/v1/stats prefix distribution from memory (closes #30)
check / check (push) Canceled after 0s

The prefix distribution was the last query on the stats path. It read one
index entry per live route on every request, so on a large database it took
the whole 4-second deadline and /api/v1/stats answered 500.

The distribution now lives next to the in-memory counts: seeded once at
startup from the same query, then kept exact by the live-route writes. A new
route whose prefix had no live route adds one at its mask length; a delete
that leaves a prefix with no live route takes one away. Each check is one
lookup on the prefix index, made only for new and removed routes, inside the
write.

Model: opus-5-5
This commit is contained in:
2026-09-29 11:05:28 +00:00
parent 6422d9fa0c
commit 539d15c0a5
5 changed files with 394 additions and 60 deletions
+66 -9
View File
@@ -6,11 +6,12 @@ import (
"sync"
)
// liveCounts holds the running row counts that the stats endpoints report. They
// are seeded once at startup from the tables and then adjusted on every write,
// so a stats read serves them from memory instead of running a COUNT(*) over
// each table. Those scans, once the database passed a few GiB, took the whole
// request timeout and made /api/v1/stats return 500 (issue 27).
// liveCounts holds the running row counts and the prefix distribution that the
// stats endpoints report. They are seeded once at startup from the tables and
// then adjusted on every write, so a stats read serves them from memory instead
// of running a query over the tables. The COUNT(*) scans (issue 27) and then
// the prefix distribution query (issue 30) each grew with the database until
// they took the whole request timeout and made /api/v1/stats return 500.
//
// A single mutex guards all fields so the stats reader takes a consistent
// snapshot at one instant and writers, which already run under the database
@@ -24,11 +25,16 @@ type liveCounts struct {
peers int
routesV4 int
routesV6 int
// The prefix distribution: for each mask length, the number of distinct
// prefixes that have at least one live route.
distributionV4 [ipv4Bits + 1]int
distributionV6 [ipv6Bits + 1]int
}
// seed sets every count to the value read from the tables at startup. It runs
// before any writer, so it needs no coordination with the adjust methods.
func (c *liveCounts) seed(asns, prefixesV4, prefixesV6, peerings, peers, routesV4, routesV6 int) {
func (c *liveCounts) seed(asns, prefixesV4, prefixesV6, peerings, peers, routesV4, routesV6 int,
distributionV4, distributionV6 []PrefixDistribution) {
c.mu.Lock()
defer c.mu.Unlock()
@@ -39,6 +45,12 @@ func (c *liveCounts) seed(asns, prefixesV4, prefixesV6, peerings, peers, routesV
c.peers = peers
c.routesV4 = routesV4
c.routesV6 = routesV6
for _, entry := range distributionV4 {
addAtMaskLength(c.distributionV4[:], entry.MaskLength, entry.Count)
}
for _, entry := range distributionV6 {
addAtMaskLength(c.distributionV6[:], entry.MaskLength, entry.Count)
}
}
// addASNs adds n to the ASN count.
@@ -79,6 +91,30 @@ func (c *liveCounts) addRoutes(v4, v6 int) {
c.mu.Unlock()
}
// addToDistribution adds n to the IPv4 and IPv6 prefix distributions once for
// each listed mask length. A write lists the mask lengths of the prefixes it
// gave their first live route with n = 1, and of the prefixes it left with no
// live route with n = -1.
func (c *liveCounts) addToDistribution(maskLengthsV4, maskLengthsV6 []int, n int) {
c.mu.Lock()
defer c.mu.Unlock()
for _, maskLength := range maskLengthsV4 {
addAtMaskLength(c.distributionV4[:], maskLength, n)
}
for _, maskLength := range maskLengthsV6 {
addAtMaskLength(c.distributionV6[:], maskLength, n)
}
}
// addAtMaskLength adds n to counts[maskLength]. A mask length the array has no
// entry for is ignored, so a malformed route cannot crash the daemon.
func addAtMaskLength(counts []int, maskLength, n int) {
if maskLength >= 0 && maskLength < len(counts) {
counts[maskLength] += n
}
}
// fill copies the counts into a Stats, including the derived totals, under a
// single read lock so the reader sees one consistent snapshot.
func (c *liveCounts) fill(s *Stats) {
@@ -94,6 +130,21 @@ func (c *liveCounts) fill(s *Stats) {
s.IPv4Routes = c.routesV4
s.IPv6Routes = c.routesV6
s.LiveRoutes = c.routesV4 + c.routesV6
s.IPv4PrefixDistribution = distributionList(c.distributionV4[:])
s.IPv6PrefixDistribution = distributionList(c.distributionV6[:])
}
// distributionList lists the mask lengths that have at least one prefix, in
// ascending order, the way the distribution query returns them.
func distributionList(counts []int) []PrefixDistribution {
var list []PrefixDistribution
for maskLength, count := range counts {
if count > 0 {
list = append(list, PrefixDistribution{MaskLength: maskLength, Count: count})
}
}
return list
}
// countRows returns the number of rows in the named table. It is used only at
@@ -108,8 +159,9 @@ func (d *Database) countRows(ctx context.Context, table string) (int, error) {
return n, nil
}
// seedCounts reads the current row counts from the tables into the in-memory
// counters. It runs once at startup, before the streamer begins writing.
// seedCounts reads the current row counts and prefix distribution from the
// tables into the in-memory counters. It runs once at startup, before the
// streamer begins writing.
func (d *Database) seedCounts(ctx context.Context) error {
asns, err := d.countRows(ctx, "asns")
if err != nil {
@@ -139,8 +191,13 @@ func (d *Database) seedCounts(ctx context.Context) error {
if err != nil {
return err
}
distributionV4, distributionV6, err := d.GetPrefixDistributionContext(ctx)
if err != nil {
return err
}
d.counts.seed(asns, prefixesV4, prefixesV6, peerings, peers, routesV4, routesV6)
d.counts.seed(asns, prefixesV4, prefixesV6, peerings, peers, routesV4, routesV6,
distributionV4, distributionV6)
return nil
}