Files
netmaker/mq/util.go
T
Abhishek Kondur 292af315dd NM-271: Scalability Improvements (#3921)
* feat(go): add user schema;

* feat(go): migrate to user schema;

* feat(go): add audit fields;

* feat(go): remove unused fields from the network model;

* feat(go): add network schema;

* feat(go): migrate to network schema;

* refactor(go): add comment to clarify migration logic;

* fix(go): test failures;

* fix(go): test failures;

* feat(go): change membership table to store memberships at all scopes;

* feat(go): add schema for access grants;

* feat(go): remove nameservers from new networks table; ensure db passed for schema functions;

* feat(go): set max conns for sqlite to 1;

* fix(go): issues updating user account status;

* NM-236: streamline operations in HA mode

* NM-236: only master pod should subscribe to updates from clients

* refactor(go): remove converters and access grants;

* refactor(go): add json tags in schema models;

* refactor(go): rename file to migrate_v1_6_0.go;

* refactor(go): add user groups and user roles tables; use schema tables;

* refactor(go): inline get and list from schema package;

* refactor(go): inline get network and list users from schema package;

* fix(go): staticcheck issues;

* fix(go): remove test not in use; fix test case;

* fix(go): validate network;

* fix(go): resolve static checks;

* fix(go): new models errors;

* fix(go): test errors;

* fix(go): handle no records;

* fix(go): add validations for user object;

* fix(go): set correct extclient status;

* fix(go): test error;

* feat(go): make schema the base package;

* feat(go): add host schema;

* feat(go): use schema host everywhere;

* feat(go): inline get host, list hosts and delete host;

* feat(go): use non-ptr value;

* feat(go): use save to upsert all fields;

* feat(go): use save to upsert all fields;

* feat(go): save turn endpoint as string;

* feat(go): check for gorm error record not found;

* fix(go): test failures;

* fix(go): update all network fields;

* fix(go): update all network fields;

* feat(go): add paginated list networks api;

* feat(go): add paginated list users api;

* feat(go): add paginated list hosts api;

* feat(go): add pagination to list groups api;

* fix(go): comment;

* fix(go): implement marshal and unmarshal text for custom types;

* fix(go): implement marshal and unmarshal json for custom types;

* fix(go): just use the old model for unmarshalling;

* fix(go): implement marshal and unmarshal json for custom types;

* NM-271:Import swap: compress/gzip replaced with github.com/klauspost/compress/gzip (2-4x faster, wire-compatible output). Added sync import.
Two sync.Pool variables (gzipWriterPool, bufferPool): reuse gzip.Writer and bytes.Buffer across calls instead of allocating fresh ones per publish.
compressPayload rewritten: pulls writer + buffer from pools, resets them, compresses at gzip.BestSpeed (level 1), copies the result out of the pooled buffer, and returns both objects to the pools.

* feat(go): remove paginated list networks api;

* feat(go): use custom paginated response object;

* NM-271: Improve server scalability under high host count

- Replace stdlib compress/gzip with klauspost/compress at BestSpeed and
  pool gzip writers and buffers via sync.Pool to eliminate compression
  as the dominant CPU hotspot.

- Debounce peer update broadcasts with a 500ms resettable window capped
  at 3s max-wait, coalescing rapid-fire PublishPeerUpdate calls into a
  single broadcast cycle.

- Cache HostPeerInfo (batch-refreshed by debounce worker) and
  HostPeerUpdate (stored as side-effect of each publish) so the pull API
  and peer_info API serve from pre-computed maps instead of triggering
  expensive per-host computations under thundering herd conditions.

- Warm both caches synchronously at startup before the first publish
  cycle so early pull requests are served instantly.

- Bound concurrent MQTT publishes to 5 via semaphore to prevent
  broker TCP buffer overflows that caused broken pipe disconnects.

- Remove manual Disconnect+SetupMQTT from ConnectionLostHandler and
  rely on the paho client's built-in AutoReconnect; add a 5s retry
  wait in publish() to ride out brief reconnection windows.

* NM-271: Reduce server CPU contention under high concurrent load

- Cache ServerSettings with atomic.Value to eliminate repeated DB reads
  on every pull request (was 32+ goroutines blocked on read lock)
- Batch UpdateNodeCheckin writes in memory, flush every 30s to reduce
  per-checkin write lock contention (was 88+ goroutines blocked)
- Enable SQLite WAL mode + busy_timeout and remove global dbMutex;
  let SQLite handle concurrency natively (reads no longer block writes)
- Move ResetFailedOverPeer/ResetAutoRelayedPeer to async in pull()
  handler since results don't affect the cached response
- Skip no-op UpsertNode writes in failover/relay reset functions
  (early return when node has no failover/relay state)
- Remove CheckHostPorts from hostUpdateFallback hot path
- Switch to pure-Go SQLite driver (glebarez/sqlite), set CGO_ENABLED=0

* fix(go): ensure default values for page and per_page are used when not passed;

* fix(go): rename v1.6.0 to v1.5.1;

* fix(go): check for gorm.ErrRecordNotFound instead of database.IsEmptyRecord;

* fix(go): use host id, not pending host id;

* NM-271: Revert pure-Go SQLite and FIPS disable to verify impact

Revert to CGO-based mattn/go-sqlite3 driver and re-enable FIPS to
isolate whether these changes are still needed now that the global
dbMutex has been removed and WAL mode is enabled. Keep WAL mode
pragma with mattn-compatible DSN format.

* feat(go): add filters to paginated apis;

* feat(go): add filters to paginated apis;

* feat(go): remove check for max username length;

* feat(go): add filters to count as well;

* feat(go): use library to check email address validity;

* feat(go): ignore pagination if params not passed;

* fix(go): pagination issues;

* fix(go): check exists before using;

* fix(go): remove debug log;

* NM-271: rm debug logs

* NM-271: check if caching is enabled

* NM-271: add server sync mq topic for HA mode

* NM-271: fix build

* NM-271: push metrics in batch to exproter over api

* NM-271: use basic auth for exporter metrics api

* fix(go): use gorm err record not found;

* NM-271: Add monitoring stack on demand

* NM-271: -m arg for install script should only add monitoring stack

* fix(go): use gorm err record not found;

* NM-271: update docker compose file for prometheus

* NM-271: update docker compose file for prometheus

* fix(go): use user principal name when creating pending user;

* fix(go): use schema package for consts;

* NM-236: rm duplicate network hook

* NM-271: add server topic to reset idp hooks on master node

* fix(go): prevent disabling superadmin user;

Co-authored-by: tenki-reviewer[bot] <262613592+tenki-reviewer[bot]@users.noreply.github.com>

* fix(go): swap is admin and is superadmin;

Co-authored-by: tenki-reviewer[bot] <262613592+tenki-reviewer[bot]@users.noreply.github.com>

* fix(go): remove dead code block;

https://github.com/gravitl/netmaker/pull/3910#discussion_r2928837937

* fix(go): incorrect message when trying to disable self;

https://github.com/gravitl/netmaker/pull/3910#discussion_r2928837934

* NM-271: fix stale peers on reset_failovered pull and add HTTP timeout to metrics exporter

Run the failover/relay reset synchronously in the pull handler so the
response reflects post-reset topology instead of serving stale cached
peers. Add a 30s timeout to the metrics exporter HTTP client to prevent
PushAllMetricsToExporter from blocking the Keepalive loop.

* NM-271: fix gzip pool corruption, MQTT topic mismatch, stale settings cache, and reduce redundant DB fetches

- Only return gzip.Writer to pool after successful Close to prevent
  silently malformed MQTT payloads from a previously errored writer.
- Fix serversync subscription to exact topic match since syncType is
  now in the message payload, not the topic path.
- Prevent zero-value ServerSettings from being cached indefinitely
  when the DB record is missing or unmarshal fails on startup.
- Return fetched hosts/nodes from RefreshHostPeerInfoCache so
  warmPeerCaches reuses them instead of querying the DB twice.
- Compute fresh HostPeerUpdate on reset_failovered pull instead of
  serving stale cache, and store result back for subsequent requests.

* NM-271: fix gzip writer pool leak, log checkin flush errors, and fix master pod ordinal parsing

- Reset gzip.Writer to io.Discard before returning to pool so errored
  writers are never leaked or silently reused with corrupt state.
- Track and log failed DB inserts in FlushNodeCheckins so operators
  have visibility when check-in timestamps are lost.
- Parse StatefulSet pod ordinal as integer instead of using HasSuffix
  to prevent netmaker-10 from being misidentified as master pod.

* NM-271: simplify masterpod logic

* fix(go): use correct header;

Co-authored-by: tenki-reviewer[bot] <262613592+tenki-reviewer[bot]@users.noreply.github.com>

* fix(go): return after error response;

Co-authored-by: tenki-reviewer[bot] <262613592+tenki-reviewer[bot]@users.noreply.github.com>

* fix(go): use correct order of params;

https://github.com/gravitl/netmaker/pull/3910#discussion_r2929593036

* fix(go): set default values for page and page size; use v2 instead of /list;

* NM-271: use host name

* Update mq/serversync.go

Co-authored-by: tenki-reviewer[bot] <262613592+tenki-reviewer[bot]@users.noreply.github.com>

* NM-271: fix duplicate serversynce case

* NM-271: streamline gw updates

* Update logic/auth.go

Co-authored-by: tenki-reviewer[bot] <262613592+tenki-reviewer[bot]@users.noreply.github.com>

* Update schema/user_roles.go

Co-authored-by: tenki-reviewer[bot] <262613592+tenki-reviewer[bot]@users.noreply.github.com>

* fix(go): syntax error;

* fix(go): set default values when page and per_page are not passed or 0;

* fix(go): use uuid.parse instead of uuid.must parse;

* fix(go): review errors;

* fix(go): review errors;

* Update controllers/user.go

Co-authored-by: tenki-reviewer[bot] <262613592+tenki-reviewer[bot]@users.noreply.github.com>

* Update controllers/user.go

Co-authored-by: tenki-reviewer[bot] <262613592+tenki-reviewer[bot]@users.noreply.github.com>

* NM-163: fix errors:

* Update db/types/options.go

Co-authored-by: tenki-reviewer[bot] <262613592+tenki-reviewer[bot]@users.noreply.github.com>

* fix(go): persist return user in event;

* Update db/types/options.go

Co-authored-by: tenki-reviewer[bot] <262613592+tenki-reviewer[bot]@users.noreply.github.com>

* NM-271: signal pull on ip changes

* NM-163: duplicate lines of code

* NM-163: fix(go): fix missing return and filter parsing in user controller

- Add missing return after error response in updateUserAccountStatus
  to prevent double-response and spurious ext-client side-effects
- Use switch statements in listUsers to skip unrecognized
  account_status and mfa_status filter values

* NM-271: signal pull req on node ip change

* fix(go): check for both min and max page size;

* NM-271: refresh node object before update

* fix(go): enclose transfer superadmin in transaction;

* fix(go): review errors;

* fix(go): remove free tier checks;

* fix(go): review fixes;

* NM-271: streamline ip pool ops

* NM-271: fix tests, set max idle conns

* NM-271: fix(go): fix data races in settings cache and peer update worker

- Use pointer type in atomic.Value for serverSettingsCache to avoid
  replacing the variable non-atomically in InvalidateServerSettingsCache
- Swap peerUpdateReplace flag before draining the channel to prevent
  a concurrent replacePeers=true from being consumed by the wrong cycle

---------

Co-authored-by: VishalDalwadi <dalwadivishal26@gmail.com>
Co-authored-by: Vishal Dalwadi <51291657+VishalDalwadi@users.noreply.github.com>
Co-authored-by: tenki-reviewer[bot] <262613592+tenki-reviewer[bot]@users.noreply.github.com>
2026-03-18 00:24:54 +05:30

235 lines
5.4 KiB
Go

package mq
import (
"bytes"
"context"
"crypto/aes"
"crypto/cipher"
"crypto/rand"
"errors"
"fmt"
"io"
"math"
"strings"
"sync"
"time"
"github.com/gravitl/netmaker/db"
"github.com/gravitl/netmaker/logic"
"github.com/gravitl/netmaker/models"
"github.com/gravitl/netmaker/netclient/ncutils"
"github.com/gravitl/netmaker/schema"
"github.com/klauspost/compress/gzip"
"golang.org/x/exp/slog"
)
func decryptMsgWithHost(host *schema.Host, msg []byte) ([]byte, error) {
if host.OS == models.OS_Types.IoT { // just pass along IoT messages
return msg, nil
}
trafficKey, trafficErr := logic.RetrievePrivateTrafficKey() // get server private key
if trafficErr != nil {
return nil, trafficErr
}
serverPrivTKey, err := ncutils.ConvertBytesToKey(trafficKey)
if err != nil {
return nil, err
}
nodePubTKey, err := ncutils.ConvertBytesToKey(host.TrafficKeyPublic)
if err != nil {
return nil, err
}
return ncutils.DeChunk(msg, nodePubTKey, serverPrivTKey)
}
func DecryptMsg(node *models.Node, msg []byte) ([]byte, error) {
if len(msg) <= 24 { // make sure message is of appropriate length
return nil, fmt.Errorf("received invalid message from broker %v", msg)
}
host := &schema.Host{ID: node.HostID}
if err := host.Get(db.WithContext(context.TODO())); err != nil {
return nil, err
}
return decryptMsgWithHost(host, msg)
}
func BatchItems[T any](items []T, batchSize int) [][]T {
if batchSize <= 0 {
return nil
}
remainderBatchSize := len(items) % batchSize
nBatches := int(math.Ceil(float64(len(items)) / float64(batchSize)))
batches := make([][]T, nBatches)
for i := range batches {
if i == nBatches-1 && remainderBatchSize > 0 {
batches[i] = make([]T, remainderBatchSize)
} else {
batches[i] = make([]T, batchSize)
}
for j := range batches[i] {
batches[i][j] = items[i*batchSize+j]
}
}
return batches
}
var gzipWriterPool = sync.Pool{
New: func() any {
w, _ := gzip.NewWriterLevel(io.Discard, gzip.BestSpeed)
return w
},
}
var bufferPool = sync.Pool{
New: func() any {
return new(bytes.Buffer)
},
}
func compressPayload(data []byte) ([]byte, error) {
buf := bufferPool.Get().(*bytes.Buffer)
buf.Reset()
defer bufferPool.Put(buf)
zw := gzipWriterPool.Get().(*gzip.Writer)
zw.Reset(buf)
defer func() {
zw.Reset(io.Discard)
gzipWriterPool.Put(zw)
}()
if _, err := zw.Write(data); err != nil {
return nil, err
}
if err := zw.Close(); err != nil {
return nil, err
}
result := make([]byte, buf.Len())
copy(result, buf.Bytes())
return result, nil
}
func encryptAESGCM(key, plaintext []byte) ([]byte, error) {
// Create AES block cipher
block, err := aes.NewCipher(key)
if err != nil {
return nil, err
}
// Create GCM (Galois/Counter Mode) cipher
aesGCM, err := cipher.NewGCM(block)
if err != nil {
return nil, err
}
// Create a random nonce
nonce := make([]byte, aesGCM.NonceSize())
if _, err := io.ReadFull(rand.Reader, nonce); err != nil {
return nil, err
}
// Encrypt the data
ciphertext := aesGCM.Seal(nonce, nonce, plaintext, nil)
return ciphertext, nil
}
func encryptMsg(host *schema.Host, msg []byte) ([]byte, error) {
if host.OS == models.OS_Types.IoT {
return msg, nil
}
// fetch server public key to be certain hasn't changed in transit
trafficKey, trafficErr := logic.RetrievePrivateTrafficKey()
if trafficErr != nil {
return nil, trafficErr
}
serverPrivKey, err := ncutils.ConvertBytesToKey(trafficKey)
if err != nil {
return nil, err
}
nodePubKey, err := ncutils.ConvertBytesToKey(host.TrafficKeyPublic)
if err != nil {
return nil, err
}
if strings.Contains(host.Version, "0.10.0") {
return ncutils.BoxEncrypt(msg, nodePubKey, serverPrivKey)
}
return ncutils.Chunk(msg, nodePubKey, serverPrivKey)
}
func publish(host *schema.Host, dest string, msg []byte) error {
var encrypted []byte
var encryptErr error
vlt, err := logic.VersionLessThan(host.Version, "v0.30.0")
if err != nil {
slog.Warn("error checking version less than", "error", err)
return err
}
if vlt {
encrypted, encryptErr = encryptMsg(host, msg)
if encryptErr != nil {
return encryptErr
}
} else {
zipped, err := compressPayload(msg)
if err != nil {
return err
}
encrypted, encryptErr = encryptAESGCM(host.TrafficKeyPublic[0:32], zipped)
if encryptErr != nil {
return encryptErr
}
}
for attempt := 0; attempt < 2; attempt++ {
if mqclient == nil || !mqclient.IsConnectionOpen() {
ok := false
for i := 0; i < 5; i++ {
time.Sleep(time.Second)
if mqclient != nil && mqclient.IsConnectionOpen() {
ok = true
break
}
}
if !ok {
return errors.New("cannot publish ... mqclient not connected")
}
}
token := mqclient.Publish(dest, 0, true, encrypted)
if token.WaitTimeout(MQ_TIMEOUT*time.Second) && token.Error() == nil {
return nil
}
if attempt == 0 {
slog.Warn("publish failed, retrying after reconnect", "dest", dest)
time.Sleep(2 * time.Second)
continue
}
if token.Error() != nil {
slog.Error("publish to mq error", "error", token.Error().Error())
return token.Error()
}
return errors.New("connection timeout")
}
return nil
}
// decodes a message queue topic and returns the embedded node.ID
func GetID(topic string) (string, error) {
parts := strings.Split(topic, "/")
count := len(parts)
if count == 1 {
return "", fmt.Errorf("invalid topic")
}
//the last part of the topic will be the node.ID
return parts[count-1], nil
}