常有人说,浏览源码是每个优良开发工程师的必经之路,然而在面对像相似 TiDB 这样简单的零碎时,源码浏览是一个十分宏大的工程。而对一些 TiDB User 来说,从本人日常遇到的问题登程,反过来浏览源码就是一个不错的切入点,因而咱们策动了《带着问题读源码》系列文章。
本文为该系列的第二篇,从一个 Power BI Desktop 在 TiDB 上体现异样的问题为例,介绍从问题的发现、定位,到通过开源社区提 issue、写 PR 解决问题的流程,从代码实现的角度来做 trouble shooting,心愿可能帮忙大家更好地理解 TiDB 源码。
首先咱们重现一下失败的场景(TiDB 5.1.1 on MacOS),建一个简略的只有一个字段的表:
CREATE TABLE test(name VARCHAR(1) PRIMARY KEY);
MySQL 上能够 TiDB 上就不能够,报错
DataSource.Error: An error happened while reading data from the provider: ‘Failed to enable constraints. One or more rows contain values violating non-null, unique, or foreign-key constraints.’
Details:
DataSourceKind=MySql
DataSourcePath=localhost:4000;test
看 general log TiDB 上最初一条跑的 SQL 是:
select COLUMN_NAME, ORDINAL_POSITION, IS_NULLABLE, DATA_TYPE, case when NUMERIC_PRECISION is null then null when DATA_TYPE in ('FLOAT', 'DOUBLE') then 2 else 10 end AS NUMERIC_PRECISION_RADIX, NUMERIC_PRECISION, NUMERIC_SCALE, CHARACTER_MAXIMUM_LENGTH, COLUMN_DEFAULT, COLUMN_COMMENT AS DESCRIPTION, COLUMN_TYPE from INFORMATION_SCHEMA.COLUMNS where table_schema = 'test' and table_name = 'test';
咱们用 tiup 启动一个 TiDB 集群,应用 tiup client 执行该命令,tiup client 也会报错:
error: mysql: sql: Scan error on column index 4, name “NUMERIC_PRECISION_RADIX”: converting NULL to int64 is unsupported
那咱们的注意力就集中在解决这条语句的问题,咱们先看 tiup client 上报的这个错意味着什么。tiup client 应用的是 golang xo/usql
库,然而在 xo/usql
库中,咱们并不能找到对应的报错信息,grep converting 关键字返回极无限且无关的内容。咱们再看 xo/usql
的 mysql driver,其中又援用到了 go-sql-driver/mysql
,下载它的代码并 grep converting,只返回了 changelog 中的一条信息,大概率报错的中央也不在这个库中。浏览一下 go-sql-driver/mysql
中的代码,发现它依赖于 database/sql
,那咱们看看 database/sql
的内容。database/sql
是 golang 的规范库,所以咱们须要下载 golang 的源码。在 golang 的 database 目录中 grep converting,很快就找到了与报错信息相符的内容:
go/src/database/sql/convert.go
case reflect.Int, reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64:
if src == nil {return fmt.Errorf("converting NULL to %s is unsupported", dv.Kind())
}
s := asString(src)
i64, err := strconv.ParseInt(s, 10, dv.Type().Bits())
if err != nil {err = strconvErr(err)
return fmt.Errorf("converting driver.Value type %T (%q) to a %s: %v", src, s, dv.Kind(), err)
}
dv.SetInt(i64)
return nil
咱们再追踪这个片段,看这里的类型是如何来的,最终咱们会回到 go-sql-driver/mysql 中:
mysql/fields.go
case fieldTypeLongLong:
if mf.flags&flagNotNULL != 0 {
if mf.flags&flagUnsigned != 0 {return scanTypeUint64}
return scanTypeInt64
}
return scanTypeNullInt
这部分的代码是在解析语句返回体中的 column definition,转换成 golang 中的类型。咱们能够应用 mysql --host 127.0.0.1 --port 4000 -u root --column-type-info
连上后查看有问题的 SQL 返回的 column metadata:
MySQL
Field 5: `NUMERIC_PRECISION_RADIX`
Catalog: `def`
Database: ``
Table: ``
Org_table: ``
Type: LONGLONG
Collation: binary (63)
Length: 3
Max_length: 0
Decimals: 0
Flags: BINARY NUM
TiDB
Field 5: `NUMERIC_PRECISION_RADIX`
Catalog: `def`
Database: ``
Table: ``
Org_table: ``
Type: LONGLONG
Collation: binary (63)
Length: 2
Max_length: 0
Decimals: 0
Flags: NOT_NULL BINARY NUM
能够很显著的看到,tiup client 报错信息中的 NUMERIC_PRECISION_RADIX
字段的 column definition 在 TiDB 上有显著的问题,该字段在 TiDB 的返回体中被标记为了 NOT_NULL,很显著这是不合理的,因为该字段显然能够是 NULL
,MySQL 的返回值也体现了这一点。所以 xo/usql
在解决返回体的时候报错了。到了这里,咱们曾经发现了 client 端为什么会报错,上面咱们就须要去寻找 TiDB 为什么会返回一个谬误的 column definition。
通过 TiDB Dev Guide 咱们能够晓得 TiDB 中一条 DQL 语句的大体执行过程,咱们从入口的 server/conn.go#clientConn.Run
往下看去,一路通过 server/conn.go#clientConn.dispatch
、server/conn.go#clientConn.handleQuery
、server/conn.go#clientConn.handleStmt
、server/driver_tidb.go#TiDBContext.ExecuteStmt
、session/session.go#session.ExecuteStmt
、executor/compiler.go#Compiler.Compile
、planner/optimize.go#Optimize
、planner/optimize.go#optimize
、planner/core/planbuilder.go#PlanBuilder.Build
、planner/core/logical_plan_builder.go#PlanBuilder.buildSelect
,在 buildSelect
中,咱们能够看到 TiDB planner 对查问语句进行的一系列解决,而后咱们就能够走到 planner/core/expression_rewriter.go#PlanBuilder.rewriteWithPreprocess
和 planner/core/expression_rewriter.go#PlanBuilder.rewriteExprNode
,在 rewriteExprNode
中,会把有问题的字段 NUMERIC_PRECISION_RADIX
进行解析,最终这条 CASE
表达式的解析会在 expression/builtin_control.go#caseWhenFunctionClass.getFunction
中,咱们终于走到了计算 CASE 表达式返回的 column definition 的中央(这依赖于遍历 compiler 解析出的 AST):
for i := 1; i < l; i += 2 {fieldTps = append(fieldTps, args[i].GetType())
decimal = mathutil.Max(decimal, args[i].GetType().Decimal)
if args[i].GetType().Flen == -1 {flen = -1} else if flen != -1 {flen = mathutil.Max(flen, args[i].GetType().Flen)
}
isBinaryStr = isBinaryStr || types.IsBinaryStr(args[i].GetType())
isBinaryFlag = isBinaryFlag || !types.IsNonBinaryStr(args[i].GetType())
}
if l%2 == 1 {fieldTps = append(fieldTps, args[l-1].GetType())
decimal = mathutil.Max(decimal, args[l-1].GetType().Decimal)
if args[l-1].GetType().Flen == -1 {flen = -1} else if flen != -1 {flen = mathutil.Max(flen, args[l-1].GetType().Flen)
}
isBinaryStr = isBinaryStr || types.IsBinaryStr(args[l-1].GetType())
isBinaryFlag = isBinaryFlag || !types.IsNonBinaryStr(args[l-1].GetType())
}
fieldTp := types.AggFieldType(fieldTps)
// Here we turn off NotNullFlag. Because if all when-clauses are false,
// the result of case-when expr is NULL.
types.SetTypeFlag(&fieldTp.Flag, mysql.NotNullFlag, false)
tp := fieldTp.EvalType()
if tp == types.ETInt {decimal = 0}
fieldTp.Decimal, fieldTp.Flen = decimal, flen
if fieldTp.EvalType().IsStringKind() && !isBinaryStr {fieldTp.Charset, fieldTp.Collate = DeriveCollationFromExprs(ctx, args...)
if fieldTp.Charset == charset.CharsetBin && fieldTp.Collate == charset.CollationBin {// When args are Json and Numerical type(eg. Int), the fieldTp is String.
// Both their charset/collation is binary, but the String need a default charset/collation.
fieldTp.Charset, fieldTp.Collate = charset.GetDefaultCharsetAndCollate()}
} else {fieldTp.Charset, fieldTp.Collate = charset.CharsetBin, charset.CollationBin}
if isBinaryFlag {fieldTp.Flag |= mysql.BinaryFlag}
// Set retType to BINARY(0) if all arguments are of type NULL.
if fieldTp.Tp == mysql.TypeNull {
fieldTp.Flen, fieldTp.Decimal = 0, types.UnspecifiedLength
types.SetBinChsClnFlag(fieldTp)
}
查看如上计算 column definition flag 的代码咱们能够发现,无论 CASE
表达式的状况是怎么样的,NOT_NULL
标记位都肯定会被设置成 false
,所以问题不呈现在这里!这个时候咱们只能沿着下面的代码门路往回看,看看下面生成的 column definition 在后续有没有被批改。终于在 server/conn.go#clientConn.handleStmt
中,发现它调用了 server/conn.go#clientConn.writeResultSet
,而后又陆续调用了 server/conn.go#clientConn.writeChunks
、server/conn.go#clientConn.writeColumnInfo
、server/column.go#ColumnInfo.Dump
和 server/column.go#dumpFlag
,在 dumpFlag 中,之前生成的 column definition flag
被批改了:
func dumpFlag(tp byte, flag uint16) uint16 {
switch tp {
case mysql.TypeSet:
return flag | uint16(mysql.SetFlag)
case mysql.TypeEnum:
return flag | uint16(mysql.EnumFlag)
default:
if mysql.HasBinaryFlag(uint(flag)) {return flag | uint16(mysql.NotNullFlag)
}
return flag
}
}
终于,咱们找到了 TiDB 返回谬误的 column definition 的起因!其实这个 bug 在 TiDB 最新版 5.2.0 中曾经被修复了:*: fix some problems related to notNullFlag by wjhuang2016 · Pull Request #27697 · pingcap/tidb。
最初,在上述浏览代码的过程中,咱们其实最好可能看到被 TiDB 解析后的 AST 是什么样子的,这样在最初遍历 AST 的过程中,才不至于摸瞎。TiDB dev guide 中有 parser 章节解说如何调试 parser,parser/quickstart.md at master · pingcap/parser 中也有样例输入生成的 AST,然而简略地输入根本没有任何作用,咱们能够应用 davecgh/go-spew
间接输入 parser 生成的 node,这样就能取得一个可被人了解的 tree:
package main
import (
"fmt"
"github.com/pingcap/parser"
"github.com/pingcap/parser/ast"
_ "github.com/pingcap/parser/test_driver"
"github.com/davecgh/go-spew/spew"
)
func parse(sql string) (*ast.StmtNode, error) {p := parser.New()
stmtNodes, _, err := p.Parse(sql, "","")
if err != nil {return nil, err}
return &stmtNodes[0], nil
}
func main() {
spew.Config.Indent = " "
astNode, err := parse("SELECT a, b FROM t")
if err != nil {fmt.Printf("parse error: %v\n", err.Error())
return
}
fmt.Printf("%s\n", spew.Sdump(*astNode))
}
(*ast.SelectStmt)(0x140001dac30)({dmlNode: (ast.dmlNode) {stmtNode: (ast.stmtNode) {node: (ast.node) {text: (string) (len=18) "SELECT a, b FROM t"
}
}
},
resultSetNode: (ast.resultSetNode) {resultFields: ([]*ast.ResultField) <nil>
},
SelectStmtOpts: (*ast.SelectStmtOpts)(0x14000115bc0)({Distinct: (bool) false,
SQLBigResult: (bool) false,
SQLBufferResult: (bool) false,
SQLCache: (bool) true,
SQLSmallResult: (bool) false,
CalcFoundRows: (bool) false,
StraightJoin: (bool) false,
Priority: (mysql.PriorityEnum) 0,
TableHints: ([]*ast.TableOptimizerHint) <nil>
}),
Distinct: (bool) false,
From: (*ast.TableRefsClause)(0x140001223c0)({node: (ast.node) {text: (string) ""
},
TableRefs: (*ast.Join)(0x14000254100)({node: (ast.node) {text: (string) ""
},
resultSetNode: (ast.resultSetNode) {resultFields: ([]*ast.ResultField) <nil>
},
Left: (*ast.TableSource)(0x14000156480)({node: (ast.node) {text: (string) ""
},
Source: (*ast.TableName)(0x1400013a370)({node: (ast.node) {text: (string) ""
},
resultSetNode: (ast.resultSetNode) {resultFields: ([]*ast.ResultField) <nil>
},
Schema: (model.CIStr) ,
Name: (model.CIStr) t,
DBInfo: (*model.DBInfo)(<nil>),
TableInfo: (*model.TableInfo)(<nil>),
IndexHints: ([]*ast.IndexHint) <nil>,
PartitionNames: ([]model.CIStr) {}}),
AsName: (model.CIStr)
}),
Right: (ast.ResultSetNode) <nil>,
Tp: (ast.JoinType) 0,
On: (*ast.OnCondition)(<nil>),
Using: ([]*ast.ColumnName) <nil>,
NaturalJoin: (bool) false,
StraightJoin: (bool) false
})
}),
Where: (ast.ExprNode) <nil>,
Fields: (*ast.FieldList)(0x14000115bf0)({node: (ast.node) {text: (string) ""
},
Fields: ([]*ast.SelectField) (len=2 cap=2) {(*ast.SelectField)(0x140001367e0)({node: (ast.node) {text: (string) (len=1) "a"
},
Offset: (int) 7,
WildCard: (*ast.WildCardField)(<nil>),
Expr: (*ast.ColumnNameExpr)(0x14000254000)({exprNode: (ast.exprNode) {node: (ast.node) {text: (string) ""
},
Type: (types.FieldType) unspecified,
flag: (uint64) 8
},
Name: (*ast.ColumnName)(0x1400017dc70)(a),
Refer: (*ast.ResultField)(<nil>)
}),
AsName: (model.CIStr) ,
Auxiliary: (bool) false
}),
(*ast.SelectField)(0x14000136840)({node: (ast.node) {text: (string) (len=1) "b"
},
Offset: (int) 10,
WildCard: (*ast.WildCardField)(<nil>),
Expr: (*ast.ColumnNameExpr)(0x14000254080)({exprNode: (ast.exprNode) {node: (ast.node) {text: (string) ""
},
Type: (types.FieldType) unspecified,
flag: (uint64) 8
},
Name: (*ast.ColumnName)(0x1400017dce0)(b),
Refer: (*ast.ResultField)(<nil>)
}),
AsName: (model.CIStr) ,
Auxiliary: (bool) false
})
}
}),
GroupBy: (*ast.GroupByClause)(<nil>),
Having: (*ast.HavingClause)(<nil>),
WindowSpecs: ([]ast.WindowSpec) <nil>,
OrderBy: (*ast.OrderByClause)(<nil>),
Limit: (*ast.Limit)(<nil>),
LockTp: (ast.SelectLockType) none,
TableHints: ([]*ast.TableOptimizerHint) <nil>,
IsAfterUnionDistinct: (bool) false,
IsInBraces: (bool) false,
QueryBlockOffset: (int) 0,
SelectIntoOpt: (*ast.SelectIntoOption)(<nil>)
})