Browse Source

network: Adaptive timeout (#7410)

* server can start when no network for local dev

* fixed superfluous response.WriteHeader call" warning

* adaptive based on last write time

* more doc

* refactoring
pull/7236/merge
Chris Lu 5 days ago
committed by GitHub
parent
commit
ba07b3e4c6
No known key found for this signature in database GPG Key ID: B5690EEEBB952194
  1. 71
      weed/util/net_timeout.go

71
weed/util/net_timeout.go

@ -1,13 +1,24 @@
package util package util
import ( import (
"github.com/seaweedfs/seaweedfs/weed/glog"
"net" "net"
"time" "time"
"github.com/seaweedfs/seaweedfs/weed/glog"
"github.com/seaweedfs/seaweedfs/weed/stats" "github.com/seaweedfs/seaweedfs/weed/stats"
) )
const (
// minThroughputBytesPerSecond defines the minimum expected throughput (4KB/s)
// Used to calculate timeout scaling based on data transferred
minThroughputBytesPerSecond = 4000
// graceTimeCapMultiplier caps the grace period for slow clients at 3x base timeout
// This prevents indefinite connections while allowing time for server-side chunk fetches
graceTimeCapMultiplier = 3
)
// Listener wraps a net.Listener, and gives a place to store the timeout // Listener wraps a net.Listener, and gives a place to store the timeout
// parameters. On Accept, it will wrap the net.Conn with our own Conn for us. // parameters. On Accept, it will wrap the net.Conn with our own Conn for us.
type Listener struct { type Listener struct {
@ -39,11 +50,28 @@ type Conn struct {
isClosed bool isClosed bool
bytesRead int64 bytesRead int64
bytesWritten int64 bytesWritten int64
lastWrite time.Time
}
// calculateBytesPerTimeout calculates the expected number of bytes that should
// be transferred during one timeout period, based on the minimum throughput.
// Returns at least 1 to prevent division by zero.
func calculateBytesPerTimeout(timeout time.Duration) int64 {
bytesPerTimeout := int64(float64(minThroughputBytesPerSecond) * timeout.Seconds())
if bytesPerTimeout <= 0 {
return 1 // Prevent division by zero
}
return bytesPerTimeout
} }
func (c *Conn) Read(b []byte) (count int, e error) { func (c *Conn) Read(b []byte) (count int, e error) {
if c.ReadTimeout != 0 { if c.ReadTimeout != 0 {
err := c.Conn.SetReadDeadline(time.Now().Add(c.ReadTimeout * time.Duration(c.bytesRead/40000+1)))
// Calculate expected bytes per timeout period based on minimum throughput (4KB/s)
// Example: with ReadTimeout=30s, bytesPerTimeout = 4000 * 30 = 120KB
// After reading 1MB: multiplier = 1,000,000/120,000 + 1 ≈ 9, deadline = 30s * 9 = 270s
bytesPerTimeout := calculateBytesPerTimeout(c.ReadTimeout)
timeoutMultiplier := time.Duration(c.bytesRead/bytesPerTimeout + 1)
err := c.Conn.SetReadDeadline(time.Now().Add(c.ReadTimeout * timeoutMultiplier))
if err != nil { if err != nil {
return 0, err return 0, err
} }
@ -58,8 +86,42 @@ func (c *Conn) Read(b []byte) (count int, e error) {
func (c *Conn) Write(b []byte) (count int, e error) { func (c *Conn) Write(b []byte) (count int, e error) {
if c.WriteTimeout != 0 { if c.WriteTimeout != 0 {
// minimum 4KB/s
err := c.Conn.SetWriteDeadline(time.Now().Add(c.WriteTimeout * time.Duration(c.bytesWritten/40000+1)))
now := time.Now()
// Calculate timeout with two components:
// 1. Base timeout scaled by cumulative data (minimum throughput of 4KB/s)
// 2. Additional grace period if there was a gap since last write (for chunk fetch delays)
// Calculate expected bytes per timeout period based on minimum throughput (4KB/s)
// Example: with WriteTimeout=30s, bytesPerTimeout = 4000 * 30 = 120KB
// After writing 1MB: multiplier = 1,000,000/120,000 + 1 ≈ 9, baseTimeout = 30s * 9 = 270s
bytesPerTimeout := calculateBytesPerTimeout(c.WriteTimeout)
timeoutMultiplier := time.Duration(c.bytesWritten/bytesPerTimeout + 1)
baseTimeout := c.WriteTimeout * timeoutMultiplier
// If it's been a while since last write, add grace time for server-side chunk fetches
// But cap it to avoid keeping slow clients connected indefinitely
//
// The comparison uses unscaled WriteTimeout intentionally: triggers grace when idle time
// exceeds base timeout, independent of throughput scaling.
if !c.lastWrite.IsZero() {
timeSinceLastWrite := now.Sub(c.lastWrite)
if timeSinceLastWrite > c.WriteTimeout {
// Add grace time capped at graceTimeCapMultiplier * scaled timeout.
// This allows total deadline up to 4x scaled timeout for server-side delays.
//
// Example: WriteTimeout=30s, 1MB written (multiplier≈9), baseTimeout=270s
// If 400s gap occurs fetching chunks: graceTime capped at 270s*3=810s
// Final deadline: 270s + 810s = 1080s (~18min) to accommodate slow storage
// But if only 50s gap: graceTime = 50s, final deadline = 270s + 50s = 320s
graceTime := timeSinceLastWrite
if graceTime > baseTimeout*graceTimeCapMultiplier {
graceTime = baseTimeout * graceTimeCapMultiplier
}
baseTimeout += graceTime
}
}
err := c.Conn.SetWriteDeadline(now.Add(baseTimeout))
if err != nil { if err != nil {
return 0, err return 0, err
} }
@ -68,6 +130,7 @@ func (c *Conn) Write(b []byte) (count int, e error) {
if e == nil { if e == nil {
stats.BytesOut(int64(count)) stats.BytesOut(int64(count))
c.bytesWritten += int64(count) c.bytesWritten += int64(count)
c.lastWrite = time.Now()
} }
return return
} }

Loading…
Cancel
Save