v0.14: various net/ fixes for large clusters (#4080)

* net.sh: Add -F to discard validator nodes that didn't bootup successfully

* Relax sanity node count when validator bootup failure is permitted

* Less sanity for testnet-demo

* net.sh: Add -F to discard validator nodes that didn't bootup successfully

ci/testnet-deploy.sh

@@ -282,39 +282,34 @@ if ! $skipStart; then
       op=start
     fi
     echo "--- net.sh $op"
+    args=("$op" -t "$tarChannelOrTag")
 
-    maybeRejectExtraNodes=
     if ! $publicNetwork; then
-      maybeRejectExtraNodes="-o rejectExtraNodes"
+      args+=(-o rejectExtraNodes)
     fi
-    maybeNoValidatorSanity=
     if [[ -n $NO_VALIDATOR_SANITY ]]; then
-      maybeNoValidatorSanity="-o noValidatorSanity"
+      args+=(-o noValidatorSanity)
     fi
-    maybeNoLedgerVerify=
     if [[ -n $NO_LEDGER_VERIFY ]]; then
-      maybeNoLedgerVerify="-o noLedgerVerify"
+      args+=(-o noLedgerVerify)
     fi
 
-    maybeReuseLedger=
     if $reuseLedger; then
-      maybeReuseLedger="-r"
+      args+=(-r)
+    fi
+
+    if ! $failOnValidatorBootupFailure; then
+      args+=(-F)
     fi
 
-    maybeUpdateManifestKeypairFile=
     # shellcheck disable=SC2154 # SOLANA_INSTALL_UPDATE_MANIFEST_KEYPAIR_x86_64_unknown_linux_gnu comes from .buildkite/env/
     if [[ -n $SOLANA_INSTALL_UPDATE_MANIFEST_KEYPAIR_x86_64_unknown_linux_gnu ]]; then
       echo "$SOLANA_INSTALL_UPDATE_MANIFEST_KEYPAIR_x86_64_unknown_linux_gnu" > update_manifest_keypair.json
-      maybeUpdateManifestKeypairFile="-i update_manifest_keypair.json"
+      args+=(-i update_manifest_keypair.json)
     fi
 
     # shellcheck disable=SC2086 # Don't want to double quote the $maybeXYZ variables
-    time net/net.sh $op -t "$tarChannelOrTag" \
-      $maybeUpdateManifestKeypairFile \
-      $maybeReuseLedger \
-      $maybeRejectExtraNodes \
-      $maybeNoValidatorSanity \
-      $maybeNoLedgerVerify
+    time net/net.sh "${args[@]}"
   ) || ok=false
 
   net/net.sh logs

ci/testnet-manager.sh

@@ -267,6 +267,7 @@ sanity() {
       ok=true
       if [[ -n $GCE_NODE_COUNT ]]; then
         NO_LEDGER_VERIFY=1 \
+        NO_VALIDATOR_SANITY=1 \
           ci/testnet-sanity.sh demo-testnet-solana-com gce "${GCE_ZONES[0]}" -f || ok=false
       else
         echo "Error: no GCE nodes"
@@ -413,6 +414,8 @@ deploy() {
       fi
 
       # shellcheck disable=SC2068
+      NO_LEDGER_VERIFY=1 \
+      NO_VALIDATOR_SANITY=1 \
         ci/testnet-deploy.sh -p demo-testnet-solana-com -C gce ${GCE_ZONE_ARGS[@]} \
           -t "$CHANNEL_OR_TAG" -n "$GCE_NODE_COUNT" -c 0 -P -u -f \
           -a demo-testnet-solana-com \
@@ -423,6 +426,8 @@ deploy() {
 
       if [[ -n $GCE_LOW_QUOTA_NODE_COUNT ]]; then
         # shellcheck disable=SC2068
+        NO_LEDGER_VERIFY=1 \
+        NO_VALIDATOR_SANITY=1 \
           ci/testnet-deploy.sh -p demo-testnet-solana-com2 -C gce ${GCE_LOW_QUOTA_ZONE_ARGS[@]} \
             -t "$CHANNEL_OR_TAG" -n "$GCE_LOW_QUOTA_NODE_COUNT" -c 0 -P -f -x \
             ${skipCreate:+-e} \

net/gce.sh

@@ -439,13 +439,15 @@ EOF
     for zone in "${zones[@]}"; do
       echo "Looking for additional fullnode instances in $zone ..."
       cloud_FindInstances "$prefix-$zone-fullnode"
-      [[ ${#instances[@]} -gt 0 ]] || {
-        echo "Unable to find additional fullnodes"
-        exit 1
-      }
-
+      if [[ ${#instances[@]} -gt 0 ]]; then
         fetchPrivateKey
         cloud_ForEachInstance recordInstanceIp "$failOnValidatorBootupFailure" fullnodeIpList
+      else
+        echo "Unable to find additional fullnodes"
+        if $failOnValidatorBootupFailure; then
+          exit 1
+        fi
+      fi
     done
   fi
 

net/net.sh

@@ -51,6 +51,7 @@ Operate a configured testnet
                                         to the bench-tps client.
 
  sanity/start/update-specific options:
+   -F                   - Discard validator nodes that didn't bootup successfully
    -o noLedgerVerify    - Skip ledger verification
    -o noValidatorSanity - Skip fullnode sanity
    -o rejectExtraNodes  - Require the exact number of nodes
@@ -80,12 +81,13 @@ numBenchTpsClients=0
 numBenchExchangeClients=0
 benchTpsExtraArgs=
 benchExchangeExtraArgs=
+failOnValidatorBootupFailure=true
 
 command=$1
 [[ -n $command ]] || usage
 shift
 
-while getopts "h?T:t:o:f:rD:i:c:" opt; do
+while getopts "h?T:t:o:f:rD:i:c:F" opt; do
   case $opt in
   h | \?)
     usage
@@ -167,6 +169,9 @@ while getopts "h?T:t:o:f:rD:i:c:" opt; do
     }
     getClientTypeAndNum
     ;;
+  F)
+    failOnValidatorBootupFailure=false
+    ;;
   *)
     usage "Error: unhandled option: $opt"
     ;;
@@ -291,6 +296,7 @@ startBootstrapLeader() {
          \"$RUST_LOG\" \
          $skipSetup \
          $leaderRotation \
+         $failOnValidatorBootupFailure \
       "
   ) >> "$logFile" 2>&1 || {
     cat "$logFile"
@@ -319,6 +325,7 @@ startNode() {
          \"$RUST_LOG\" \
          $skipSetup \
          $leaderRotation \
+         $failOnValidatorBootupFailure \
       "
   ) >> "$logFile" 2>&1 &
   declare pid=$!
@@ -475,9 +482,13 @@ start() {
     declare ok=true
     wait "$pid" || ok=false
     if ! $ok; then
+      echo "+++ fullnode failed to start"
       cat "$netLogDir/fullnode-$pid.log"
-      echo ^^^ +++
+      if $failOnValidatorBootupFailure; then
         exit 1
+      else
+        echo "Failure is non-fatal"
+      fi
     fi
   done
 

net/remote/remote-node.sh

@@ -12,6 +12,7 @@ numNodes="$5"
 RUST_LOG="$6"
 skipSetup="$7"
 leaderRotation="$8"
+failOnValidatorBootupFailure="$9"
 set +x
 export RUST_LOG
 
@@ -35,12 +36,14 @@ missing() {
 [[ -n $numNodes ]]      || missing numNodes
 [[ -n $skipSetup ]]     || missing skipSetup
 [[ -n $leaderRotation ]] || missing leaderRotation
+[[ -n $failOnValidatorBootupFailure ]] || missing failOnValidatorBootupFailure
 
 cat > deployConfig <<EOF
 deployMethod="$deployMethod"
 entrypointIp="$entrypointIp"
 numNodes="$numNodes"
 leaderRotation=$leaderRotation
+failOnValidatorBootupFailure=$failOnValidatorBootupFailure
 EOF
 
 source net/common.sh

net/remote/remote-sanity.sh

@@ -9,6 +9,7 @@ cd "$(dirname "$0")"/../..
 deployMethod=
 entrypointIp=
 numNodes=
+failOnValidatorBootupFailure=
 
 [[ -r deployConfig ]] || {
   echo deployConfig missing
@@ -26,6 +27,7 @@ missing() {
 [[ -n $entrypointIp ]]   || missing entrypointIp
 [[ -n $numNodes ]]       || missing numNodes
 [[ -n $leaderRotation ]] || missing leaderRotation
+[[ -n $failOnValidatorBootupFailure ]] || missing failOnValidatorBootupFailure
 
 ledgerVerify=true
 validatorSanity=true
@@ -79,7 +81,17 @@ local|tar)
   exit 1
 esac
 
-echo "+++ $entrypointIp: node count ($numNodes expected)"
+if $failOnValidatorBootupFailure; then
+  numSanityNodes=1
+  if $rejectExtraNodes; then
+    echo "rejectExtraNodes cannot be used with failOnValidatorBootupFailure"
+    exit 1
+  fi
+else
+  numSanityNodes="$numNodes"
+fi
+
+echo "+++ $entrypointIp: node count ($numSanityNodes expected)"
 (
   set -x
   $solana_keygen -o "$client_id"
@@ -90,7 +102,7 @@ echo "+++ $entrypointIp: node count ($numNodes expected)"
   fi
 
   timeout 2m $solana_gossip --network "$entrypointIp:8001" \
-    spy --$nodeArg "$numNodes" \
+    spy --$nodeArg "$numSanityNodes" \
 )
 
 echo "--- RPC API: getTransactionCount"

net/scripts/gce-provider.sh

@@ -34,8 +34,10 @@ __cloud_FindInstances() {
     instances+=("$name:$publicIp:$privateIp:$zone")
   done < <(gcloud compute instances list \
              --filter "$filter" \
-             --format 'value(name,networkInterfaces[0].accessConfigs[0].natIP,networkInterfaces[0].networkIP,status,zone)')
+             --format 'value(name,networkInterfaces[0].accessConfigs[0].natIP,networkInterfaces[0].networkIP,status,zone)' \
+           | grep RUNNING)
 }
+
 #
 # cloud_FindInstances [namePrefix]
 #