VictoriaMetrics/lib/logstorage/pipe_math.go
Aliaksandr Valialkin 246c339e3d
lib/logstorage: read timestamps column when it is really needed during query execution
Previously timestamps column was read unconditionally on every query.
This could significantly slow down queries, which do not need reading this column
like in https://github.com/VictoriaMetrics/VictoriaMetrics/issues/7070 .
2024-09-25 19:18:37 +02:00

935 lines
20 KiB
Go

package logstorage
import (
"fmt"
"math"
"strings"
"unsafe"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/decimal"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/slicesutil"
)
// pipeMath processes '| math ...' pipe.
//
// See https://docs.victoriametrics.com/victorialogs/logsql/#math-pipe
type pipeMath struct {
entries []*mathEntry
}
type mathEntry struct {
// The calculated expr result is stored in resultField.
resultField string
// expr is the expression to calculate.
expr *mathExpr
}
type mathExpr struct {
// if isConst is set, then the given mathExpr returns the given constValue.
isConst bool
constValue float64
// constValueStr is the original string representation of constValue.
//
// It is used in String() method for returning the original representation of the given constValue.
constValueStr string
// if fieldName isn't empty, then the given mathExpr fetches numeric values from the given fieldName.
fieldName string
// args are args for the given mathExpr.
args []*mathExpr
// op is the operation name (aka function name) for the given mathExpr.
op string
// f is the function for calculating results for the given mathExpr.
f mathFunc
// whether the mathExpr was wrapped in parens.
wrappedInParens bool
}
// mathFunc must fill result with calculated results based on the given args.
type mathFunc func(result []float64, args [][]float64)
func (pm *pipeMath) String() string {
s := "math"
a := make([]string, len(pm.entries))
for i, e := range pm.entries {
a[i] = e.String()
}
s += " " + strings.Join(a, ", ")
return s
}
func (pm *pipeMath) canLiveTail() bool {
return true
}
func (me *mathEntry) String() string {
s := me.expr.String()
if isMathBinaryOp(me.expr.op) {
s = "(" + s + ")"
}
s += " as " + quoteTokenIfNeeded(me.resultField)
return s
}
func (me *mathExpr) String() string {
if me.isConst {
return me.constValueStr
}
if me.fieldName != "" {
return quoteTokenIfNeeded(me.fieldName)
}
args := me.args
if isMathBinaryOp(me.op) {
opPriority := getMathBinaryOpPriority(me.op)
left := me.args[0]
right := me.args[1]
leftStr := left.String()
rightStr := right.String()
if isMathBinaryOp(left.op) && getMathBinaryOpPriority(left.op) > opPriority {
leftStr = "(" + leftStr + ")"
}
if isMathBinaryOp(right.op) && getMathBinaryOpPriority(right.op) > opPriority {
rightStr = "(" + rightStr + ")"
}
return fmt.Sprintf("%s %s %s", leftStr, me.op, rightStr)
}
if me.op == "unary_minus" {
argStr := args[0].String()
if isMathBinaryOp(args[0].op) {
argStr = "(" + argStr + ")"
}
return "-" + argStr
}
a := make([]string, len(args))
for i, arg := range args {
a[i] = arg.String()
}
argsStr := strings.Join(a, ", ")
return fmt.Sprintf("%s(%s)", me.op, argsStr)
}
func isMathBinaryOp(op string) bool {
_, ok := mathBinaryOps[op]
return ok
}
func getMathBinaryOpPriority(op string) int {
bo, ok := mathBinaryOps[op]
if !ok {
logger.Panicf("BUG: unexpected binary op: %q", op)
}
return bo.priority
}
func getMathFuncForBinaryOp(op string) (mathFunc, error) {
bo, ok := mathBinaryOps[op]
if !ok {
return nil, fmt.Errorf("unsupported binary operation: %q", op)
}
return bo.f, nil
}
var mathBinaryOps = map[string]mathBinaryOp{
"^": {
priority: 1,
f: mathFuncPow,
},
"*": {
priority: 2,
f: mathFuncMul,
},
"/": {
priority: 2,
f: mathFuncDiv,
},
"%": {
priority: 2,
f: mathFuncMod,
},
"+": {
priority: 3,
f: mathFuncPlus,
},
"-": {
priority: 3,
f: mathFuncMinus,
},
"&": {
priority: 4,
f: mathFuncAnd,
},
"xor": {
priority: 5,
f: mathFuncXor,
},
"or": {
priority: 6,
f: mathFuncOr,
},
"default": {
priority: 10,
f: mathFuncDefault,
},
}
type mathBinaryOp struct {
priority int
f mathFunc
}
func (pm *pipeMath) updateNeededFields(neededFields, unneededFields fieldsSet) {
for i := len(pm.entries) - 1; i >= 0; i-- {
e := pm.entries[i]
if neededFields.contains("*") {
if !unneededFields.contains(e.resultField) {
unneededFields.add(e.resultField)
fs := newFieldsSet()
e.expr.updateNeededFields(fs)
unneededFields.removeFields(fs.getAll())
}
} else {
if neededFields.contains(e.resultField) {
neededFields.remove(e.resultField)
e.expr.updateNeededFields(neededFields)
}
}
}
}
func (me *mathExpr) updateNeededFields(neededFields fieldsSet) {
if me.isConst {
return
}
if me.fieldName != "" {
neededFields.add(me.fieldName)
return
}
for _, arg := range me.args {
arg.updateNeededFields(neededFields)
}
}
func (pm *pipeMath) optimize() {
// nothing to do
}
func (pm *pipeMath) hasFilterInWithQuery() bool {
return false
}
func (pm *pipeMath) initFilterInValues(_ map[string][]string, _ getFieldValuesFunc) (pipe, error) {
return pm, nil
}
func (pm *pipeMath) newPipeProcessor(workersCount int, _ <-chan struct{}, _ func(), ppNext pipeProcessor) pipeProcessor {
pmp := &pipeMathProcessor{
pm: pm,
ppNext: ppNext,
shards: make([]pipeMathProcessorShard, workersCount),
}
return pmp
}
type pipeMathProcessor struct {
pm *pipeMath
ppNext pipeProcessor
shards []pipeMathProcessorShard
}
type pipeMathProcessorShard struct {
pipeMathProcessorShardNopad
// The padding prevents false sharing on widespread platforms with 128 mod (cache line size) = 0 .
_ [128 - unsafe.Sizeof(pipeMathProcessorShardNopad{})%128]byte
}
type pipeMathProcessorShardNopad struct {
// a holds all the data for rcs.
a arena
// rcs is used for storing calculated results before they are written to ppNext.
rcs []resultColumn
// rs is storage for temporary results
rs [][]float64
// rsBuf is backing storage for rs slices
rsBuf []float64
}
func (shard *pipeMathProcessorShard) executeMathEntry(e *mathEntry, rc *resultColumn, br *blockResult) {
clear(shard.rs)
shard.rs = shard.rs[:0]
shard.rsBuf = shard.rsBuf[:0]
shard.executeExpr(e.expr, br)
r := shard.rs[0]
b := shard.a.b
for _, f := range r {
bLen := len(b)
b = marshalFloat64String(b, f)
v := bytesutil.ToUnsafeString(b[bLen:])
rc.addValue(v)
}
shard.a.b = b
}
func (shard *pipeMathProcessorShard) executeExpr(me *mathExpr, br *blockResult) {
rIdx := len(shard.rs)
shard.rs = slicesutil.SetLength(shard.rs, len(shard.rs)+1)
shard.rsBuf = slicesutil.SetLength(shard.rsBuf, len(shard.rsBuf)+br.rowsLen)
shard.rs[rIdx] = shard.rsBuf[len(shard.rsBuf)-br.rowsLen:]
if me.isConst {
r := shard.rs[rIdx]
for i := 0; i < br.rowsLen; i++ {
r[i] = me.constValue
}
return
}
if me.fieldName != "" {
c := br.getColumnByName(me.fieldName)
values := c.getValues(br)
r := shard.rs[rIdx]
var f float64
for i, v := range values {
if i == 0 || v != values[i-1] {
f = parseMathNumber(v)
}
r[i] = f
}
return
}
rsBufLen := len(shard.rsBuf)
for _, arg := range me.args {
shard.executeExpr(arg, br)
}
result := shard.rs[rIdx]
args := shard.rs[rIdx+1:]
me.f(result, args)
shard.rs = shard.rs[:rIdx+1]
shard.rsBuf = shard.rsBuf[:rsBufLen]
}
func (pmp *pipeMathProcessor) writeBlock(workerID uint, br *blockResult) {
if br.rowsLen == 0 {
return
}
shard := &pmp.shards[workerID]
entries := pmp.pm.entries
shard.rcs = slicesutil.SetLength(shard.rcs, len(entries))
rcs := shard.rcs
for i, e := range entries {
rc := &rcs[i]
rc.name = e.resultField
shard.executeMathEntry(e, rc, br)
br.addResultColumn(rc)
}
pmp.ppNext.writeBlock(workerID, br)
for i := range rcs {
rcs[i].resetValues()
}
shard.a.reset()
}
func (pmp *pipeMathProcessor) flush() error {
return nil
}
func parsePipeMath(lex *lexer) (*pipeMath, error) {
if !lex.isKeyword("math", "eval") {
return nil, fmt.Errorf("unexpected token: %q; want 'math' or 'eval'", lex.token)
}
lex.nextToken()
var mes []*mathEntry
for {
me, err := parseMathEntry(lex)
if err != nil {
return nil, err
}
mes = append(mes, me)
switch {
case lex.isKeyword(","):
lex.nextToken()
case lex.isKeyword("|", ")", ""):
if len(mes) == 0 {
return nil, fmt.Errorf("missing 'math' expressions")
}
pm := &pipeMath{
entries: mes,
}
return pm, nil
default:
return nil, fmt.Errorf("unexpected token after 'math' expression [%s]: %q; expecting ',', '|' or ')'", mes[len(mes)-1], lex.token)
}
}
}
func parseMathEntry(lex *lexer) (*mathEntry, error) {
me, err := parseMathExpr(lex)
if err != nil {
return nil, err
}
resultField := ""
if lex.isKeyword(",", "|", ")", "") {
resultField = me.String()
} else {
if lex.isKeyword("as") {
// skip optional 'as'
lex.nextToken()
}
fieldName, err := parseFieldName(lex)
if err != nil {
return nil, fmt.Errorf("cannot parse result name for [%s]: %w", me, err)
}
resultField = fieldName
}
e := &mathEntry{
resultField: resultField,
expr: me,
}
return e, nil
}
func parseMathExpr(lex *lexer) (*mathExpr, error) {
// parse left operand
left, err := parseMathExprOperand(lex)
if err != nil {
return nil, err
}
for {
if !isMathBinaryOp(lex.token) {
// There is no right operand
return left, nil
}
// parse operator
op := lex.token
lex.nextToken()
f, err := getMathFuncForBinaryOp(op)
if err != nil {
return nil, fmt.Errorf("cannot parse operator after [%s]: %w", left, err)
}
// parse right operand
right, err := parseMathExprOperand(lex)
if err != nil {
return nil, fmt.Errorf("cannot parse operand after [%s %s]: %w", left, op, err)
}
me := &mathExpr{
args: []*mathExpr{left, right},
op: op,
f: f,
}
// balance operands according to their priority
if !left.wrappedInParens && isMathBinaryOp(left.op) && getMathBinaryOpPriority(left.op) > getMathBinaryOpPriority(op) {
me.args[0] = left.args[1]
left.args[1] = me
me = left
}
left = me
}
}
func parseMathExprInParens(lex *lexer) (*mathExpr, error) {
if !lex.isKeyword("(") {
return nil, fmt.Errorf("missing '('")
}
lex.nextToken()
me, err := parseMathExpr(lex)
if err != nil {
return nil, err
}
me.wrappedInParens = true
if !lex.isKeyword(")") {
return nil, fmt.Errorf("missing ')'; got %q instead", lex.token)
}
lex.nextToken()
return me, nil
}
func parseMathExprOperand(lex *lexer) (*mathExpr, error) {
if lex.isKeyword("(") {
return parseMathExprInParens(lex)
}
switch {
case lex.isKeyword("abs"):
return parseMathExprAbs(lex)
case lex.isKeyword("exp"):
return parseMathExprExp(lex)
case lex.isKeyword("ln"):
return parseMathExprLn(lex)
case lex.isKeyword("max"):
return parseMathExprMax(lex)
case lex.isKeyword("min"):
return parseMathExprMin(lex)
case lex.isKeyword("round"):
return parseMathExprRound(lex)
case lex.isKeyword("ceil"):
return parseMathExprCeil(lex)
case lex.isKeyword("floor"):
return parseMathExprFloor(lex)
case lex.isKeyword("-"):
return parseMathExprUnaryMinus(lex)
case lex.isKeyword("+"):
// just skip unary plus
lex.nextToken()
return parseMathExprOperand(lex)
case isNumberPrefix(lex.token):
return parseMathExprConstNumber(lex)
default:
return parseMathExprFieldName(lex)
}
}
func parseMathExprAbs(lex *lexer) (*mathExpr, error) {
me, err := parseMathExprGenericFunc(lex, "abs", mathFuncAbs)
if err != nil {
return nil, err
}
if len(me.args) != 1 {
return nil, fmt.Errorf("'abs' function accepts only one arg; got %d args: [%s]", len(me.args), me)
}
return me, nil
}
func parseMathExprExp(lex *lexer) (*mathExpr, error) {
me, err := parseMathExprGenericFunc(lex, "exp", mathFuncExp)
if err != nil {
return nil, err
}
if len(me.args) != 1 {
return nil, fmt.Errorf("'exp' function accepts only one arg; got %d args: [%s]", len(me.args), me)
}
return me, nil
}
func parseMathExprLn(lex *lexer) (*mathExpr, error) {
me, err := parseMathExprGenericFunc(lex, "ln", mathFuncLn)
if err != nil {
return nil, err
}
if len(me.args) != 1 {
return nil, fmt.Errorf("'ln' function accepts only one arg; got %d args: [%s]", len(me.args), me)
}
return me, nil
}
func parseMathExprMax(lex *lexer) (*mathExpr, error) {
me, err := parseMathExprGenericFunc(lex, "max", mathFuncMax)
if err != nil {
return nil, err
}
if len(me.args) < 2 {
return nil, fmt.Errorf("'max' function needs at least 2 args; got %d args: [%s]", len(me.args), me)
}
return me, nil
}
func parseMathExprMin(lex *lexer) (*mathExpr, error) {
me, err := parseMathExprGenericFunc(lex, "min", mathFuncMin)
if err != nil {
return nil, err
}
if len(me.args) < 2 {
return nil, fmt.Errorf("'min' function needs at least 2 args; got %d args: [%s]", len(me.args), me)
}
return me, nil
}
func parseMathExprRound(lex *lexer) (*mathExpr, error) {
me, err := parseMathExprGenericFunc(lex, "round", mathFuncRound)
if err != nil {
return nil, err
}
if len(me.args) != 1 && len(me.args) != 2 {
return nil, fmt.Errorf("'round' function needs 1 or 2 args; got %d args: [%s]", len(me.args), me)
}
return me, nil
}
func parseMathExprCeil(lex *lexer) (*mathExpr, error) {
me, err := parseMathExprGenericFunc(lex, "ceil", mathFuncCeil)
if err != nil {
return nil, err
}
if len(me.args) != 1 {
return nil, fmt.Errorf("'ceil' function needs one arg; got %d args: [%s]", len(me.args), me)
}
return me, nil
}
func parseMathExprFloor(lex *lexer) (*mathExpr, error) {
me, err := parseMathExprGenericFunc(lex, "floor", mathFuncFloor)
if err != nil {
return nil, err
}
if len(me.args) != 1 {
return nil, fmt.Errorf("'floor' function needs one arg; got %d args: [%s]", len(me.args), me)
}
return me, nil
}
func parseMathExprGenericFunc(lex *lexer, funcName string, f mathFunc) (*mathExpr, error) {
if !lex.isKeyword(funcName) {
return nil, fmt.Errorf("missing %q keyword", funcName)
}
lex.nextToken()
args, err := parseMathFuncArgs(lex)
if err != nil {
return nil, fmt.Errorf("cannot parse args for %q function: %w", funcName, err)
}
if len(args) == 0 {
return nil, fmt.Errorf("%q function needs at least one org", funcName)
}
me := &mathExpr{
args: args,
op: funcName,
f: f,
}
return me, nil
}
func parseMathFuncArgs(lex *lexer) ([]*mathExpr, error) {
if !lex.isKeyword("(") {
return nil, fmt.Errorf("missing '('")
}
lex.nextToken()
var args []*mathExpr
for {
if lex.isKeyword(")") {
lex.nextToken()
return args, nil
}
me, err := parseMathExpr(lex)
if err != nil {
return nil, err
}
args = append(args, me)
switch {
case lex.isKeyword(")"):
case lex.isKeyword(","):
lex.nextToken()
default:
return nil, fmt.Errorf("unexpected token after [%s]: %q; want ',' or ')'", me, lex.token)
}
}
}
func parseMathExprUnaryMinus(lex *lexer) (*mathExpr, error) {
if !lex.isKeyword("-") {
return nil, fmt.Errorf("missing '-'")
}
lex.nextToken()
expr, err := parseMathExprOperand(lex)
if err != nil {
return nil, err
}
me := &mathExpr{
args: []*mathExpr{expr},
op: "unary_minus",
f: mathFuncUnaryMinus,
}
return me, nil
}
func parseMathExprConstNumber(lex *lexer) (*mathExpr, error) {
if !isNumberPrefix(lex.token) {
return nil, fmt.Errorf("cannot parse number from %q", lex.token)
}
numStr, err := getCompoundMathToken(lex)
if err != nil {
return nil, fmt.Errorf("cannot parse number: %w", err)
}
f := parseMathNumber(numStr)
if math.IsNaN(f) {
return nil, fmt.Errorf("cannot parse number from %q", numStr)
}
me := &mathExpr{
isConst: true,
constValue: f,
constValueStr: numStr,
}
return me, nil
}
func parseMathExprFieldName(lex *lexer) (*mathExpr, error) {
fieldName, err := getCompoundMathToken(lex)
if err != nil {
return nil, err
}
fieldName = getCanonicalColumnName(fieldName)
me := &mathExpr{
fieldName: fieldName,
}
return me, nil
}
func getCompoundMathToken(lex *lexer) (string, error) {
stopTokens := []string{"=", "+", "-", "*", "/", "%", "^", ",", ")", "|", "!", ""}
if lex.isKeyword(stopTokens...) {
return "", fmt.Errorf("compound token cannot start with '%s'", lex.token)
}
s := lex.token
rawS := lex.rawToken
lex.nextToken()
suffix := ""
for !lex.isSkippedSpace && !lex.isKeyword(stopTokens...) {
s += lex.token
lex.nextToken()
}
if suffix == "" {
return s, nil
}
return rawS + suffix, nil
}
func mathFuncAnd(result []float64, args [][]float64) {
a := args[0]
b := args[1]
for i := range result {
if math.IsNaN(a[i]) || math.IsNaN(b[i]) {
result[i] = nan
} else {
result[i] = float64(uint64(a[i]) & uint64(b[i]))
}
}
}
func mathFuncOr(result []float64, args [][]float64) {
a := args[0]
b := args[1]
for i := range result {
if math.IsNaN(a[i]) || math.IsNaN(b[i]) {
result[i] = nan
} else {
result[i] = float64(uint64(a[i]) | uint64(b[i]))
}
}
}
func mathFuncXor(result []float64, args [][]float64) {
a := args[0]
b := args[1]
for i := range result {
if math.IsNaN(a[i]) || math.IsNaN(b[i]) {
result[i] = nan
} else {
result[i] = float64(uint64(a[i]) ^ uint64(b[i]))
}
}
}
func mathFuncPlus(result []float64, args [][]float64) {
a := args[0]
b := args[1]
for i := range result {
result[i] = a[i] + b[i]
}
}
func mathFuncMinus(result []float64, args [][]float64) {
a := args[0]
b := args[1]
for i := range result {
result[i] = a[i] - b[i]
}
}
func mathFuncMul(result []float64, args [][]float64) {
a := args[0]
b := args[1]
for i := range result {
result[i] = a[i] * b[i]
}
}
func mathFuncDiv(result []float64, args [][]float64) {
a := args[0]
b := args[1]
for i := range result {
result[i] = a[i] / b[i]
}
}
func mathFuncMod(result []float64, args [][]float64) {
a := args[0]
b := args[1]
for i := range result {
result[i] = math.Mod(a[i], b[i])
}
}
func mathFuncPow(result []float64, args [][]float64) {
a := args[0]
b := args[1]
for i := range result {
result[i] = math.Pow(a[i], b[i])
}
}
func mathFuncDefault(result []float64, args [][]float64) {
values := args[0]
defaultValues := args[1]
for i := range result {
f := values[i]
if math.IsNaN(f) {
f = defaultValues[i]
}
result[i] = f
}
}
func mathFuncAbs(result []float64, args [][]float64) {
arg := args[0]
for i := range result {
result[i] = math.Abs(arg[i])
}
}
func mathFuncExp(result []float64, args [][]float64) {
arg := args[0]
for i := range result {
result[i] = math.Exp(arg[i])
}
}
func mathFuncLn(result []float64, args [][]float64) {
arg := args[0]
for i := range result {
result[i] = math.Log(arg[i])
}
}
func mathFuncUnaryMinus(result []float64, args [][]float64) {
arg := args[0]
for i := range result {
result[i] = -arg[i]
}
}
func mathFuncMax(result []float64, args [][]float64) {
for i := range result {
f := nan
for _, arg := range args {
if math.IsNaN(f) || arg[i] > f {
f = arg[i]
}
}
result[i] = f
}
}
func mathFuncMin(result []float64, args [][]float64) {
for i := range result {
f := nan
for _, arg := range args {
if math.IsNaN(f) || arg[i] < f {
f = arg[i]
}
}
result[i] = f
}
}
func mathFuncCeil(result []float64, args [][]float64) {
arg := args[0]
for i := range result {
result[i] = math.Ceil(arg[i])
}
}
func mathFuncFloor(result []float64, args [][]float64) {
arg := args[0]
for i := range result {
result[i] = math.Floor(arg[i])
}
}
func mathFuncRound(result []float64, args [][]float64) {
arg := args[0]
if len(args) == 1 {
// Round to integer
for i := range result {
result[i] = math.Round(arg[i])
}
return
}
// Round to nearest
nearest := args[1]
var f float64
for i := range result {
if i == 0 || arg[i-1] != arg[i] || nearest[i-1] != nearest[i] {
f = round(arg[i], nearest[i])
}
result[i] = f
}
}
func round(f, nearest float64) float64 {
_, e := decimal.FromFloat(nearest)
p10 := math.Pow10(int(-e))
f += 0.5 * math.Copysign(nearest, f)
f -= math.Mod(f, nearest)
f, _ = math.Modf(f * p10)
return f / p10
}
func parseMathNumber(s string) float64 {
f, ok := tryParseNumber(s)
if ok {
return f
}
nsecs, ok := TryParseTimestampRFC3339Nano(s)
if ok {
return float64(nsecs)
}
ipNum, ok := tryParseIPv4(s)
if ok {
return float64(ipNum)
}
return nan
}