storj/satellite/satellitedb/nodeselection.go

312 lines
8.5 KiB
Go
Raw Normal View History

// Copyright (C) 2019 Storj Labs, Inc.
// See LICENSE for copying information.
package satellitedb
import (
"context"
"database/sql"
"fmt"
"strings"
"time"
"github.com/zeebo/errs"
"storj.io/common/pb"
"storj.io/common/storj"
"storj.io/private/dbutil/pgutil"
"storj.io/private/version"
"storj.io/storj/satellite/overlay"
)
func (cache *overlaycache) SelectStorageNodes(ctx context.Context, totalNeededNodes, newNodeCount int, criteria *overlay.NodeCriteria) (nodes []*overlay.SelectedNode, err error) {
defer mon.Task()(&ctx)(&err)
if totalNeededNodes == 0 {
return nil, nil
}
if newNodeCount > totalNeededNodes {
return nil, Error.New("requested new node count can't exceed requested total node count")
}
needNewNodes := newNodeCount
needReputableNodes := totalNeededNodes - needNewNodes
receivedNewNodes := 0
receivedNodeNetworks := make(map[string]struct{})
excludedIDs := append([]storj.NodeID{}, criteria.ExcludedIDs...)
excludedNetworks := append([]string{}, criteria.ExcludedNetworks...)
for i := 0; i < 3; i++ {
reputableNodes, newNodes, err := cache.selectStorageNodesOnce(ctx, needReputableNodes, needNewNodes, criteria, excludedIDs, excludedNetworks)
if err != nil {
return nil, err
}
for _, node := range newNodes {
// checking for last net collision among reputable and new nodes since we can't check within the query
if _, ok := receivedNodeNetworks[node.LastNet]; ok {
continue
}
excludedIDs = append(excludedIDs, node.ID)
excludedNetworks = append(excludedNetworks, node.LastNet)
nodes = append(nodes, node)
needNewNodes--
receivedNewNodes++
satellite/overlay: configurable meaning of last_net Up to now, we have been implementing the DistinctIP preference with code in two places: 1. On check-in, the last_net is determined by taking the /24 or /64 (in ResolveIPAndNetwork()) and we store it with the node record. 2. On node selection, a preference parameter defines whether to return results that are distinct on last_net. It can be observed that we have never yet had the need to switch from DistinctIP to !DistinctIP, or from !DistinctIP to DistinctIP, on the same satellite, and we will probably never need to do so in an automated way. It can also be observed that this arrangement makes tests more complicated, because we often have to arrange for test nodes to have IP addresses in different /24 networks (a particular pain on macOS). Those two considerations, plus some pending work on the repair framework that will make repair take last_net into consideration, motivate this change. With this change, in the #2 place, we will _always_ return results that are distinct on last_net. We implement the DistinctIP preference, then, by making the #1 place (ResolveIPAndNetwork()) more flexible. When DistinctIP is enabled, last_net will be calculated as it was before. But when DistinctIP is _off_, last_net can be the same as address (IP and port). That will effectively implement !DistinctIP because every record will have a distinct last_net already. As a side effect, this flexibility will allow us to change the rules about last_net construction arbitrarily. We can do tests where last_net is set to the source IP, or to a /30 prefix, or a /16 prefix, etc., and be able to exercise the production logic without requiring a virtual network bridge. This change should be safe to make without any migration code, because all known production satellite deployments use DistinctIP, and the associated last_net values will not change for them. They will only change for satellites with !DistinctIP, which are mostly test deployments that can be recreated trivially. For those satellites which are both permanent and !DistinctIP, node selection will suddenly start acting as though DistinctIP is enabled, until the operator runs a single SQL update "UPDATE nodes SET last_net = last_ip_port". That can be done either before or after deploying software with this change. I also assert that this will not hurt performance for production deployments. It's true that adding the distinct requirement to node selection makes things a little slower, but the distinct requirement is already present for all production deployments, and they will see no change. Refs: https://github.com/storj/storj/issues/5391 Change-Id: I0e7e92498c3da768df5b4d5fb213dcd2d4862924
2023-02-28 22:57:39 +00:00
receivedNodeNetworks[node.LastNet] = struct{}{}
}
for _, node := range reputableNodes {
if _, ok := receivedNodeNetworks[node.LastNet]; ok {
continue
}
excludedIDs = append(excludedIDs, node.ID)
excludedNetworks = append(excludedNetworks, node.LastNet)
nodes = append(nodes, node)
needReputableNodes--
satellite/overlay: configurable meaning of last_net Up to now, we have been implementing the DistinctIP preference with code in two places: 1. On check-in, the last_net is determined by taking the /24 or /64 (in ResolveIPAndNetwork()) and we store it with the node record. 2. On node selection, a preference parameter defines whether to return results that are distinct on last_net. It can be observed that we have never yet had the need to switch from DistinctIP to !DistinctIP, or from !DistinctIP to DistinctIP, on the same satellite, and we will probably never need to do so in an automated way. It can also be observed that this arrangement makes tests more complicated, because we often have to arrange for test nodes to have IP addresses in different /24 networks (a particular pain on macOS). Those two considerations, plus some pending work on the repair framework that will make repair take last_net into consideration, motivate this change. With this change, in the #2 place, we will _always_ return results that are distinct on last_net. We implement the DistinctIP preference, then, by making the #1 place (ResolveIPAndNetwork()) more flexible. When DistinctIP is enabled, last_net will be calculated as it was before. But when DistinctIP is _off_, last_net can be the same as address (IP and port). That will effectively implement !DistinctIP because every record will have a distinct last_net already. As a side effect, this flexibility will allow us to change the rules about last_net construction arbitrarily. We can do tests where last_net is set to the source IP, or to a /30 prefix, or a /16 prefix, etc., and be able to exercise the production logic without requiring a virtual network bridge. This change should be safe to make without any migration code, because all known production satellite deployments use DistinctIP, and the associated last_net values will not change for them. They will only change for satellites with !DistinctIP, which are mostly test deployments that can be recreated trivially. For those satellites which are both permanent and !DistinctIP, node selection will suddenly start acting as though DistinctIP is enabled, until the operator runs a single SQL update "UPDATE nodes SET last_net = last_ip_port". That can be done either before or after deploying software with this change. I also assert that this will not hurt performance for production deployments. It's true that adding the distinct requirement to node selection makes things a little slower, but the distinct requirement is already present for all production deployments, and they will see no change. Refs: https://github.com/storj/storj/issues/5391 Change-Id: I0e7e92498c3da768df5b4d5fb213dcd2d4862924
2023-02-28 22:57:39 +00:00
receivedNodeNetworks[node.LastNet] = struct{}{}
}
// when we did not find new nodes, then return all as reputable
if needNewNodes > 0 && receivedNewNodes == 0 {
needReputableNodes += needNewNodes
needNewNodes = 0
}
if needReputableNodes <= 0 && needNewNodes <= 0 {
break
}
}
return nodes, nil
}
func (cache *overlaycache) selectStorageNodesOnce(ctx context.Context, reputableNodeCount, newNodeCount int, criteria *overlay.NodeCriteria, excludedIDs []storj.NodeID, excludedNetworks []string) (reputableNodes, newNodes []*overlay.SelectedNode, err error) {
defer mon.Task()(&ctx)(&err)
newNodesCondition, err := nodeSelectionCondition(ctx, criteria, excludedIDs, excludedNetworks, true)
if err != nil {
return nil, nil, err
}
reputableNodesCondition, err := nodeSelectionCondition(ctx, criteria, excludedIDs, excludedNetworks, false)
if err != nil {
return nil, nil, err
}
var reputableNodeQuery, newNodeQuery partialQuery
// Note: the true/false at the end of each selection string indicates if the selection is for new nodes or not.
// Later, the flag allows us to distinguish if a node is new when scanning the db rows.
satellite/overlay: configurable meaning of last_net Up to now, we have been implementing the DistinctIP preference with code in two places: 1. On check-in, the last_net is determined by taking the /24 or /64 (in ResolveIPAndNetwork()) and we store it with the node record. 2. On node selection, a preference parameter defines whether to return results that are distinct on last_net. It can be observed that we have never yet had the need to switch from DistinctIP to !DistinctIP, or from !DistinctIP to DistinctIP, on the same satellite, and we will probably never need to do so in an automated way. It can also be observed that this arrangement makes tests more complicated, because we often have to arrange for test nodes to have IP addresses in different /24 networks (a particular pain on macOS). Those two considerations, plus some pending work on the repair framework that will make repair take last_net into consideration, motivate this change. With this change, in the #2 place, we will _always_ return results that are distinct on last_net. We implement the DistinctIP preference, then, by making the #1 place (ResolveIPAndNetwork()) more flexible. When DistinctIP is enabled, last_net will be calculated as it was before. But when DistinctIP is _off_, last_net can be the same as address (IP and port). That will effectively implement !DistinctIP because every record will have a distinct last_net already. As a side effect, this flexibility will allow us to change the rules about last_net construction arbitrarily. We can do tests where last_net is set to the source IP, or to a /30 prefix, or a /16 prefix, etc., and be able to exercise the production logic without requiring a virtual network bridge. This change should be safe to make without any migration code, because all known production satellite deployments use DistinctIP, and the associated last_net values will not change for them. They will only change for satellites with !DistinctIP, which are mostly test deployments that can be recreated trivially. For those satellites which are both permanent and !DistinctIP, node selection will suddenly start acting as though DistinctIP is enabled, until the operator runs a single SQL update "UPDATE nodes SET last_net = last_ip_port". That can be done either before or after deploying software with this change. I also assert that this will not hurt performance for production deployments. It's true that adding the distinct requirement to node selection makes things a little slower, but the distinct requirement is already present for all production deployments, and they will see no change. Refs: https://github.com/storj/storj/issues/5391 Change-Id: I0e7e92498c3da768df5b4d5fb213dcd2d4862924
2023-02-28 22:57:39 +00:00
reputableNodeQuery = partialQuery{
selection: `SELECT DISTINCT ON (last_net) last_net, id, address, last_ip_port, noise_proto, noise_public_key, debounce_limit, features, false FROM nodes`,
satellite/overlay: configurable meaning of last_net Up to now, we have been implementing the DistinctIP preference with code in two places: 1. On check-in, the last_net is determined by taking the /24 or /64 (in ResolveIPAndNetwork()) and we store it with the node record. 2. On node selection, a preference parameter defines whether to return results that are distinct on last_net. It can be observed that we have never yet had the need to switch from DistinctIP to !DistinctIP, or from !DistinctIP to DistinctIP, on the same satellite, and we will probably never need to do so in an automated way. It can also be observed that this arrangement makes tests more complicated, because we often have to arrange for test nodes to have IP addresses in different /24 networks (a particular pain on macOS). Those two considerations, plus some pending work on the repair framework that will make repair take last_net into consideration, motivate this change. With this change, in the #2 place, we will _always_ return results that are distinct on last_net. We implement the DistinctIP preference, then, by making the #1 place (ResolveIPAndNetwork()) more flexible. When DistinctIP is enabled, last_net will be calculated as it was before. But when DistinctIP is _off_, last_net can be the same as address (IP and port). That will effectively implement !DistinctIP because every record will have a distinct last_net already. As a side effect, this flexibility will allow us to change the rules about last_net construction arbitrarily. We can do tests where last_net is set to the source IP, or to a /30 prefix, or a /16 prefix, etc., and be able to exercise the production logic without requiring a virtual network bridge. This change should be safe to make without any migration code, because all known production satellite deployments use DistinctIP, and the associated last_net values will not change for them. They will only change for satellites with !DistinctIP, which are mostly test deployments that can be recreated trivially. For those satellites which are both permanent and !DistinctIP, node selection will suddenly start acting as though DistinctIP is enabled, until the operator runs a single SQL update "UPDATE nodes SET last_net = last_ip_port". That can be done either before or after deploying software with this change. I also assert that this will not hurt performance for production deployments. It's true that adding the distinct requirement to node selection makes things a little slower, but the distinct requirement is already present for all production deployments, and they will see no change. Refs: https://github.com/storj/storj/issues/5391 Change-Id: I0e7e92498c3da768df5b4d5fb213dcd2d4862924
2023-02-28 22:57:39 +00:00
condition: reputableNodesCondition,
distinct: true,
limit: reputableNodeCount,
orderBy: "last_net",
}
newNodeQuery = partialQuery{
selection: `SELECT DISTINCT ON (last_net) last_net, id, address, last_ip_port, noise_proto, noise_public_key, debounce_limit, features, true FROM nodes`,
satellite/overlay: configurable meaning of last_net Up to now, we have been implementing the DistinctIP preference with code in two places: 1. On check-in, the last_net is determined by taking the /24 or /64 (in ResolveIPAndNetwork()) and we store it with the node record. 2. On node selection, a preference parameter defines whether to return results that are distinct on last_net. It can be observed that we have never yet had the need to switch from DistinctIP to !DistinctIP, or from !DistinctIP to DistinctIP, on the same satellite, and we will probably never need to do so in an automated way. It can also be observed that this arrangement makes tests more complicated, because we often have to arrange for test nodes to have IP addresses in different /24 networks (a particular pain on macOS). Those two considerations, plus some pending work on the repair framework that will make repair take last_net into consideration, motivate this change. With this change, in the #2 place, we will _always_ return results that are distinct on last_net. We implement the DistinctIP preference, then, by making the #1 place (ResolveIPAndNetwork()) more flexible. When DistinctIP is enabled, last_net will be calculated as it was before. But when DistinctIP is _off_, last_net can be the same as address (IP and port). That will effectively implement !DistinctIP because every record will have a distinct last_net already. As a side effect, this flexibility will allow us to change the rules about last_net construction arbitrarily. We can do tests where last_net is set to the source IP, or to a /30 prefix, or a /16 prefix, etc., and be able to exercise the production logic without requiring a virtual network bridge. This change should be safe to make without any migration code, because all known production satellite deployments use DistinctIP, and the associated last_net values will not change for them. They will only change for satellites with !DistinctIP, which are mostly test deployments that can be recreated trivially. For those satellites which are both permanent and !DistinctIP, node selection will suddenly start acting as though DistinctIP is enabled, until the operator runs a single SQL update "UPDATE nodes SET last_net = last_ip_port". That can be done either before or after deploying software with this change. I also assert that this will not hurt performance for production deployments. It's true that adding the distinct requirement to node selection makes things a little slower, but the distinct requirement is already present for all production deployments, and they will see no change. Refs: https://github.com/storj/storj/issues/5391 Change-Id: I0e7e92498c3da768df5b4d5fb213dcd2d4862924
2023-02-28 22:57:39 +00:00
condition: newNodesCondition,
distinct: true,
limit: newNodeCount,
orderBy: "last_net",
}
query := unionAll(newNodeQuery, reputableNodeQuery)
query.query = cache.db.impl.WrapAsOfSystemInterval(query.query, criteria.AsOfSystemInterval)
rows, err := cache.db.Query(ctx, cache.db.Rebind(query.query), query.args...)
if err != nil {
return nil, nil, Error.Wrap(err)
}
defer func() { err = errs.Combine(err, rows.Close()) }()
for rows.Next() {
var node overlay.SelectedNode
node.Address = &pb.NodeAddress{}
var lastIPPort sql.NullString
var isNew bool
var noise noiseScanner
err = rows.Scan(&node.LastNet, &node.ID, &node.Address.Address, &node.LastIPPort, &noise.Proto, &noise.PublicKey, &node.Address.DebounceLimit, &node.Address.Features, &isNew)
if err != nil {
return nil, nil, err
}
if lastIPPort.Valid {
node.LastIPPort = lastIPPort.String
}
node.Address.NoiseInfo = noise.Convert()
if isNew {
newNodes = append(newNodes, &node)
} else {
reputableNodes = append(reputableNodes, &node)
}
if len(newNodes) >= newNodeCount && len(reputableNodes) >= reputableNodeCount {
break
}
}
return reputableNodes, newNodes, Error.Wrap(rows.Err())
}
// nodeSelectionCondition creates a condition with arguments that corresponds to the arguments.
func nodeSelectionCondition(ctx context.Context, criteria *overlay.NodeCriteria, excludedIDs []storj.NodeID, excludedNetworks []string, isNewNodeQuery bool) (condition, error) {
var conds conditions
conds.add(`disqualified IS NULL`)
conds.add(`unknown_audit_suspended IS NULL`)
conds.add(`offline_suspended IS NULL`)
conds.add(`exit_initiated_at IS NULL`)
conds.add(`type = ?`, int(pb.NodeType_STORAGE))
conds.add(`free_disk >= ?`, criteria.FreeDisk)
conds.add(`last_contact_success > ?`, time.Now().UTC().Add(-criteria.OnlineWindow))
if isNewNodeQuery {
conds.add(
`vetted_at IS NULL`,
)
} else {
conds.add(
`vetted_at is NOT NULL`,
)
}
if criteria.MinimumVersion != "" {
v, err := version.NewSemVer(criteria.MinimumVersion)
if err != nil {
return condition{}, Error.New("invalid node selection criteria version: %v", err)
}
conds.add(
`(major > ? OR (major = ? AND (minor > ? OR (minor = ? AND patch >= ?)))) AND release`,
v.Major, v.Major, v.Minor, v.Minor, v.Patch,
)
}
if len(excludedIDs) > 0 {
conds.add(
`NOT (id = ANY(?::bytea[]))`,
pgutil.NodeIDArray(excludedIDs),
)
}
satellite/overlay: configurable meaning of last_net Up to now, we have been implementing the DistinctIP preference with code in two places: 1. On check-in, the last_net is determined by taking the /24 or /64 (in ResolveIPAndNetwork()) and we store it with the node record. 2. On node selection, a preference parameter defines whether to return results that are distinct on last_net. It can be observed that we have never yet had the need to switch from DistinctIP to !DistinctIP, or from !DistinctIP to DistinctIP, on the same satellite, and we will probably never need to do so in an automated way. It can also be observed that this arrangement makes tests more complicated, because we often have to arrange for test nodes to have IP addresses in different /24 networks (a particular pain on macOS). Those two considerations, plus some pending work on the repair framework that will make repair take last_net into consideration, motivate this change. With this change, in the #2 place, we will _always_ return results that are distinct on last_net. We implement the DistinctIP preference, then, by making the #1 place (ResolveIPAndNetwork()) more flexible. When DistinctIP is enabled, last_net will be calculated as it was before. But when DistinctIP is _off_, last_net can be the same as address (IP and port). That will effectively implement !DistinctIP because every record will have a distinct last_net already. As a side effect, this flexibility will allow us to change the rules about last_net construction arbitrarily. We can do tests where last_net is set to the source IP, or to a /30 prefix, or a /16 prefix, etc., and be able to exercise the production logic without requiring a virtual network bridge. This change should be safe to make without any migration code, because all known production satellite deployments use DistinctIP, and the associated last_net values will not change for them. They will only change for satellites with !DistinctIP, which are mostly test deployments that can be recreated trivially. For those satellites which are both permanent and !DistinctIP, node selection will suddenly start acting as though DistinctIP is enabled, until the operator runs a single SQL update "UPDATE nodes SET last_net = last_ip_port". That can be done either before or after deploying software with this change. I also assert that this will not hurt performance for production deployments. It's true that adding the distinct requirement to node selection makes things a little slower, but the distinct requirement is already present for all production deployments, and they will see no change. Refs: https://github.com/storj/storj/issues/5391 Change-Id: I0e7e92498c3da768df5b4d5fb213dcd2d4862924
2023-02-28 22:57:39 +00:00
if len(excludedNetworks) > 0 {
conds.add(
`NOT (last_net = ANY(?::text[]))`,
pgutil.TextArray(excludedNetworks),
)
}
satellite/overlay: configurable meaning of last_net Up to now, we have been implementing the DistinctIP preference with code in two places: 1. On check-in, the last_net is determined by taking the /24 or /64 (in ResolveIPAndNetwork()) and we store it with the node record. 2. On node selection, a preference parameter defines whether to return results that are distinct on last_net. It can be observed that we have never yet had the need to switch from DistinctIP to !DistinctIP, or from !DistinctIP to DistinctIP, on the same satellite, and we will probably never need to do so in an automated way. It can also be observed that this arrangement makes tests more complicated, because we often have to arrange for test nodes to have IP addresses in different /24 networks (a particular pain on macOS). Those two considerations, plus some pending work on the repair framework that will make repair take last_net into consideration, motivate this change. With this change, in the #2 place, we will _always_ return results that are distinct on last_net. We implement the DistinctIP preference, then, by making the #1 place (ResolveIPAndNetwork()) more flexible. When DistinctIP is enabled, last_net will be calculated as it was before. But when DistinctIP is _off_, last_net can be the same as address (IP and port). That will effectively implement !DistinctIP because every record will have a distinct last_net already. As a side effect, this flexibility will allow us to change the rules about last_net construction arbitrarily. We can do tests where last_net is set to the source IP, or to a /30 prefix, or a /16 prefix, etc., and be able to exercise the production logic without requiring a virtual network bridge. This change should be safe to make without any migration code, because all known production satellite deployments use DistinctIP, and the associated last_net values will not change for them. They will only change for satellites with !DistinctIP, which are mostly test deployments that can be recreated trivially. For those satellites which are both permanent and !DistinctIP, node selection will suddenly start acting as though DistinctIP is enabled, until the operator runs a single SQL update "UPDATE nodes SET last_net = last_ip_port". That can be done either before or after deploying software with this change. I also assert that this will not hurt performance for production deployments. It's true that adding the distinct requirement to node selection makes things a little slower, but the distinct requirement is already present for all production deployments, and they will see no change. Refs: https://github.com/storj/storj/issues/5391 Change-Id: I0e7e92498c3da768df5b4d5fb213dcd2d4862924
2023-02-28 22:57:39 +00:00
conds.add(`last_net <> ''`)
return conds.combine(), nil
}
// partialQuery corresponds to a query.
//
// distinct=false
//
// $selection WHERE $condition ORDER BY $orderBy, RANDOM() LIMIT $limit
//
// distinct=true
//
// SELECT * FROM ($selection WHERE $condition ORDER BY $orderBy, RANDOM()) filtered ORDER BY RANDOM() LIMIT $limit
type partialQuery struct {
selection string
condition condition
distinct bool
orderBy string
limit int
}
// isEmpty returns whether the result for the query is definitely empty.
func (partial partialQuery) isEmpty() bool {
return partial.limit == 0
}
// asQuery combines partialQuery parameters into a single select query.
func (partial partialQuery) asQuery() query {
var q strings.Builder
var args []interface{}
if partial.distinct {
// For distinct queries we need to redo randomized ordering.
fmt.Fprintf(&q, "SELECT * FROM (")
}
fmt.Fprint(&q, partial.selection, " WHERE ", partial.condition.query)
args = append(args, partial.condition.args...)
if partial.orderBy != "" {
fmt.Fprintf(&q, " ORDER BY %s, RANDOM() ", partial.orderBy)
} else {
fmt.Fprint(&q, " ORDER BY RANDOM() ")
}
if !partial.distinct {
fmt.Fprint(&q, " LIMIT ? ")
args = append(args, partial.limit)
} else {
fmt.Fprint(&q, ") filtered ORDER BY RANDOM() LIMIT ?")
args = append(args, partial.limit)
}
return query{query: q.String(), args: args}
}
// unionAll combines multiple partial queries into a single query.
func unionAll(partials ...partialQuery) query {
var queries []string
var args []interface{}
for _, partial := range partials {
if partial.isEmpty() {
continue
}
q := partial.asQuery()
queries = append(queries, q.query)
args = append(args, q.args...)
}
if len(queries) == 0 {
return query{}
}
if len(queries) == 1 {
return query{query: queries[0], args: args}
}
return query{
query: "(" + strings.Join(queries, ") UNION ALL (") + ")",
args: args,
}
}
type condition struct {
query string
args []interface{}
}
type conditions []condition
func (conds *conditions) add(q string, args ...interface{}) {
*conds = append(*conds, condition{query: q, args: args})
}
func (conds conditions) combine() condition {
var qs []string
var args []interface{}
for _, c := range conds {
qs = append(qs, c.query)
args = append(args, c.args...)
}
return condition{query: " " + strings.Join(qs, " AND ") + " ", args: args}
}
type query struct {
query string
args []interface{}
}