State storage refactor (#1839)

* Hash-deduplicated state storage (and migrations) for PostgreSQL and SQLite

* Refactor droomserver database setup for migrations

* Fix conflict statements

* Update migration names

* Set a boundary for old to new block/snapshot IDs so we don't rewrite them more than once accidentally

* Create sequence if not exists

* Fix boundary queries

* Fix boundary queries

* Use Query

* Break out queries a bit

* More sequence tweaks

* Query parameters are not playing the game

* Injection escaping may not work for CREATE SEQUENCE after all

* Fix snapshot sequence name

* Use boundaried IDs in SQLite too

* Use IFNULL for SQLite

* Use COALESCE in PostgreSQL

* Review comments @Kegsay
This commit is contained in:
Neil Alexander 2021-04-26 13:25:57 +01:00 committed by GitHub
parent d6e9b7b307
commit 5ce1fe80de
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
39 changed files with 1076 additions and 554 deletions

View file

@ -18,6 +18,7 @@ package sqlite3
import (
"context"
"database/sql"
"encoding/json"
"fmt"
"sort"
"strings"
@ -32,228 +33,113 @@ import (
const stateDataSchema = `
CREATE TABLE IF NOT EXISTS roomserver_state_block (
state_block_nid INTEGER NOT NULL,
event_type_nid INTEGER NOT NULL,
event_state_key_nid INTEGER NOT NULL,
event_nid INTEGER NOT NULL,
UNIQUE (state_block_nid, event_type_nid, event_state_key_nid)
-- The state snapshot NID that identifies this snapshot.
state_block_nid INTEGER PRIMARY KEY AUTOINCREMENT,
-- The hash of the state block, which is used to enforce uniqueness. The hash is
-- generated in Dendrite and passed through to the database, as a btree index over
-- this column is cheap and fits within the maximum index size.
state_block_hash BLOB UNIQUE,
-- The event NIDs contained within the state block, encoded as JSON.
event_nids TEXT NOT NULL DEFAULT '[]'
);
`
const insertStateDataSQL = "" +
"INSERT INTO roomserver_state_block (state_block_nid, event_type_nid, event_state_key_nid, event_nid)" +
" VALUES ($1, $2, $3, $4)"
const selectNextStateBlockNIDSQL = `
SELECT IFNULL(MAX(state_block_nid), 0) + 1 FROM roomserver_state_block
// Insert a new state block. If we conflict on the hash column then
// we must perform an update so that the RETURNING statement returns the
// ID of the row that we conflicted with, so that we can then refer to
// the original block.
const insertStateDataSQL = `
INSERT INTO roomserver_state_block (state_block_hash, event_nids)
VALUES ($1, $2)
ON CONFLICT (state_block_hash) DO UPDATE SET event_nids=$2
RETURNING state_block_nid
`
// Bulk state lookup by numeric state block ID.
// Sort by the state_block_nid, event_type_nid, event_state_key_nid
// This means that all the entries for a given state_block_nid will appear
// together in the list and those entries will sorted by event_type_nid
// and event_state_key_nid. This property makes it easier to merge two
// state data blocks together.
const bulkSelectStateBlockEntriesSQL = "" +
"SELECT state_block_nid, event_type_nid, event_state_key_nid, event_nid" +
" FROM roomserver_state_block WHERE state_block_nid IN ($1)" +
" ORDER BY state_block_nid, event_type_nid, event_state_key_nid"
// Bulk state lookup by numeric state block ID.
// Filters the rows in each block to the requested types and state keys.
// We would like to restrict to particular type state key pairs but we are
// restricted by the query language to pull the cross product of a list
// of types and a list state_keys. So we have to filter the result in the
// application to restrict it to the list of event types and state keys we
// actually wanted.
const bulkSelectFilteredStateBlockEntriesSQL = "" +
"SELECT state_block_nid, event_type_nid, event_state_key_nid, event_nid" +
" FROM roomserver_state_block WHERE state_block_nid IN ($1)" +
" AND event_type_nid IN ($2) AND event_state_key_nid IN ($3)" +
" ORDER BY state_block_nid, event_type_nid, event_state_key_nid"
"SELECT state_block_nid, event_nids" +
" FROM roomserver_state_block WHERE state_block_nid IN ($1)"
type stateBlockStatements struct {
db *sql.DB
insertStateDataStmt *sql.Stmt
selectNextStateBlockNIDStmt *sql.Stmt
bulkSelectStateBlockEntriesStmt *sql.Stmt
bulkSelectFilteredStateBlockEntriesStmt *sql.Stmt
db *sql.DB
insertStateDataStmt *sql.Stmt
bulkSelectStateBlockEntriesStmt *sql.Stmt
}
func NewSqliteStateBlockTable(db *sql.DB) (tables.StateBlock, error) {
func createStateBlockTable(db *sql.DB) error {
_, err := db.Exec(stateDataSchema)
return err
}
func prepareStateBlockTable(db *sql.DB) (tables.StateBlock, error) {
s := &stateBlockStatements{
db: db,
}
_, err := db.Exec(stateDataSchema)
if err != nil {
return nil, err
}
return s, shared.StatementList{
{&s.insertStateDataStmt, insertStateDataSQL},
{&s.selectNextStateBlockNIDStmt, selectNextStateBlockNIDSQL},
{&s.bulkSelectStateBlockEntriesStmt, bulkSelectStateBlockEntriesSQL},
{&s.bulkSelectFilteredStateBlockEntriesStmt, bulkSelectFilteredStateBlockEntriesSQL},
}.Prepare(db)
}
func (s *stateBlockStatements) BulkInsertStateData(
ctx context.Context, txn *sql.Tx,
entries []types.StateEntry,
) (types.StateBlockNID, error) {
if len(entries) == 0 {
return 0, nil
ctx context.Context,
txn *sql.Tx,
entries types.StateEntries,
) (id types.StateBlockNID, err error) {
entries = entries[:util.SortAndUnique(entries)]
var nids types.EventNIDs
for _, e := range entries {
nids = append(nids, e.EventNID)
}
var stateBlockNID types.StateBlockNID
err := sqlutil.TxStmt(txn, s.selectNextStateBlockNIDStmt).QueryRowContext(ctx).Scan(&stateBlockNID)
js, err := json.Marshal(nids)
if err != nil {
return 0, err
return 0, fmt.Errorf("json.Marshal: %w", err)
}
for _, entry := range entries {
_, err = sqlutil.TxStmt(txn, s.insertStateDataStmt).ExecContext(
ctx,
int64(stateBlockNID),
int64(entry.EventTypeNID),
int64(entry.EventStateKeyNID),
int64(entry.EventNID),
)
if err != nil {
return 0, err
}
}
return stateBlockNID, err
err = s.insertStateDataStmt.QueryRowContext(
ctx, nids.Hash(), js,
).Scan(&id)
return
}
func (s *stateBlockStatements) BulkSelectStateBlockEntries(
ctx context.Context, stateBlockNIDs []types.StateBlockNID,
) ([]types.StateEntryList, error) {
nids := make([]interface{}, len(stateBlockNIDs))
for k, v := range stateBlockNIDs {
nids[k] = v
ctx context.Context, stateBlockNIDs types.StateBlockNIDs,
) ([][]types.EventNID, error) {
intfs := make([]interface{}, len(stateBlockNIDs))
for i := range stateBlockNIDs {
intfs[i] = int64(stateBlockNIDs[i])
}
selectOrig := strings.Replace(bulkSelectStateBlockEntriesSQL, "($1)", sqlutil.QueryVariadic(len(nids)), 1)
selectOrig := strings.Replace(bulkSelectStateBlockEntriesSQL, "($1)", sqlutil.QueryVariadic(len(intfs)), 1)
selectStmt, err := s.db.Prepare(selectOrig)
if err != nil {
return nil, err
}
rows, err := selectStmt.QueryContext(ctx, nids...)
rows, err := selectStmt.QueryContext(ctx, intfs...)
if err != nil {
return nil, err
}
defer internal.CloseAndLogIfError(ctx, rows, "bulkSelectStateBlockEntries: rows.close() failed")
results := make([]types.StateEntryList, len(stateBlockNIDs))
// current is a pointer to the StateEntryList to append the state entries to.
var current *types.StateEntryList
results := make([][]types.EventNID, len(stateBlockNIDs))
i := 0
for rows.Next() {
var (
stateBlockNID int64
eventTypeNID int64
eventStateKeyNID int64
eventNID int64
entry types.StateEntry
)
if err := rows.Scan(
&stateBlockNID, &eventTypeNID, &eventStateKeyNID, &eventNID,
); err != nil {
for ; rows.Next(); i++ {
var stateBlockNID types.StateBlockNID
var result json.RawMessage
if err = rows.Scan(&stateBlockNID, &result); err != nil {
return nil, err
}
entry.EventTypeNID = types.EventTypeNID(eventTypeNID)
entry.EventStateKeyNID = types.EventStateKeyNID(eventStateKeyNID)
entry.EventNID = types.EventNID(eventNID)
if current == nil || types.StateBlockNID(stateBlockNID) != current.StateBlockNID {
// The state entry row is for a different state data block to the current one.
// So we start appending to the next entry in the list.
current = &results[i]
current.StateBlockNID = types.StateBlockNID(stateBlockNID)
i++
r := []types.EventNID{}
if err = json.Unmarshal(result, &r); err != nil {
return nil, fmt.Errorf("json.Unmarshal: %w", err)
}
current.StateEntries = append(current.StateEntries, entry)
results[i] = r
}
if i != len(nids) {
return nil, fmt.Errorf("storage: state data NIDs missing from the database (%d != %d)", i, len(nids))
}
return results, nil
}
func (s *stateBlockStatements) BulkSelectFilteredStateBlockEntries(
ctx context.Context,
stateBlockNIDs []types.StateBlockNID,
stateKeyTuples []types.StateKeyTuple,
) ([]types.StateEntryList, error) {
tuples := stateKeyTupleSorter(stateKeyTuples)
// Sort the tuples so that we can run binary search against them as we filter the rows returned by the db.
sort.Sort(tuples)
eventTypeNIDArray, eventStateKeyNIDArray := tuples.typesAndStateKeysAsArrays()
sqlStatement := strings.Replace(bulkSelectFilteredStateBlockEntriesSQL, "($1)", sqlutil.QueryVariadic(len(stateBlockNIDs)), 1)
sqlStatement = strings.Replace(sqlStatement, "($2)", sqlutil.QueryVariadicOffset(len(eventTypeNIDArray), len(stateBlockNIDs)), 1)
sqlStatement = strings.Replace(sqlStatement, "($3)", sqlutil.QueryVariadicOffset(len(eventStateKeyNIDArray), len(stateBlockNIDs)+len(eventTypeNIDArray)), 1)
var params []interface{}
for _, val := range stateBlockNIDs {
params = append(params, int64(val))
}
for _, val := range eventTypeNIDArray {
params = append(params, val)
}
for _, val := range eventStateKeyNIDArray {
params = append(params, val)
}
rows, err := s.db.QueryContext(
ctx,
sqlStatement,
params...,
)
if err != nil {
if err = rows.Err(); err != nil {
return nil, err
}
defer internal.CloseAndLogIfError(ctx, rows, "bulkSelectFilteredStateBlockEntries: rows.close() failed")
var results []types.StateEntryList
var current types.StateEntryList
for rows.Next() {
var (
stateBlockNID int64
eventTypeNID int64
eventStateKeyNID int64
eventNID int64
entry types.StateEntry
)
if err := rows.Scan(
&stateBlockNID, &eventTypeNID, &eventStateKeyNID, &eventNID,
); err != nil {
return nil, err
}
entry.EventTypeNID = types.EventTypeNID(eventTypeNID)
entry.EventStateKeyNID = types.EventStateKeyNID(eventStateKeyNID)
entry.EventNID = types.EventNID(eventNID)
// We can use binary search here because we sorted the tuples earlier
if !tuples.contains(entry.StateKeyTuple) {
// The select will return the cross product of types and state keys.
// So we need to check if type of the entry is in the list.
continue
}
if types.StateBlockNID(stateBlockNID) != current.StateBlockNID {
// The state entry row is for a different state data block to the current one.
// So we append the current entry to the results and start adding to a new one.
// The first time through the loop current will be empty.
if current.StateEntries != nil {
results = append(results, current)
}
current = types.StateEntryList{StateBlockNID: types.StateBlockNID(stateBlockNID)}
}
current.StateEntries = append(current.StateEntries, entry)
if i != len(stateBlockNIDs) {
return nil, fmt.Errorf("storage: state data NIDs missing from the database (%d != %d)", len(results), len(stateBlockNIDs))
}
// Add the last entry to the list if it is not empty.
if current.StateEntries != nil {
results = append(results, current)
}
return results, nil
return results, err
}
type stateKeyTupleSorter []types.StateKeyTuple