update to latest bstore, with a bugfix for queries with multiple orders that were partially handled by an index

causing returned order to be incorrect.
was triggered by new code i'm working on.
This commit is contained in:
Mechiel Lukkien 2024-03-30 09:39:18 +01:00
parent 54b24931c9
commit d34dd8aae6
No known key found for this signature in database
10 changed files with 235 additions and 135 deletions

2
go.mod
View file

@ -5,7 +5,7 @@ go 1.21
require ( require (
github.com/mjl-/adns v0.0.0-20240309142737-2a1aacf346dc github.com/mjl-/adns v0.0.0-20240309142737-2a1aacf346dc
github.com/mjl-/autocert v0.0.0-20231214125928-31b7400acb05 github.com/mjl-/autocert v0.0.0-20231214125928-31b7400acb05
github.com/mjl-/bstore v0.0.4 github.com/mjl-/bstore v0.0.5
github.com/mjl-/sconf v0.0.6 github.com/mjl-/sconf v0.0.6
github.com/mjl-/sherpa v0.6.7 github.com/mjl-/sherpa v0.6.7
github.com/mjl-/sherpadoc v0.0.12 github.com/mjl-/sherpadoc v0.0.12

4
go.sum
View file

@ -28,8 +28,8 @@ github.com/mjl-/adns v0.0.0-20240309142737-2a1aacf346dc h1:ghTx3KsrO0hSJW0bCFCGw
github.com/mjl-/adns v0.0.0-20240309142737-2a1aacf346dc/go.mod h1:v47qUMJnipnmDTRGaHwpCwzE6oypa5K33mUvBfzZBn8= github.com/mjl-/adns v0.0.0-20240309142737-2a1aacf346dc/go.mod h1:v47qUMJnipnmDTRGaHwpCwzE6oypa5K33mUvBfzZBn8=
github.com/mjl-/autocert v0.0.0-20231214125928-31b7400acb05 h1:s6ay4bh4tmpPLdxjyeWG45mcwHfEluBMuGPkqxHWUJ4= github.com/mjl-/autocert v0.0.0-20231214125928-31b7400acb05 h1:s6ay4bh4tmpPLdxjyeWG45mcwHfEluBMuGPkqxHWUJ4=
github.com/mjl-/autocert v0.0.0-20231214125928-31b7400acb05/go.mod h1:taMFU86abMxKLPV4Bynhv8enbYmS67b8LG80qZv2Qus= github.com/mjl-/autocert v0.0.0-20231214125928-31b7400acb05/go.mod h1:taMFU86abMxKLPV4Bynhv8enbYmS67b8LG80qZv2Qus=
github.com/mjl-/bstore v0.0.4 h1:q+R1oAr8+E9yf9q+zxkVjQ18VFqD/E9KmGVoe4FIOBA= github.com/mjl-/bstore v0.0.5 h1:Cx+LWEBnFBsqSxZNMxeVujkfc0kG10lUJaAU4vWSRHo=
github.com/mjl-/bstore v0.0.4/go.mod h1:/cD25FNBaDfvL/plFRxI3Ba3E+wcB0XVOS8nJDqndg0= github.com/mjl-/bstore v0.0.5/go.mod h1:/cD25FNBaDfvL/plFRxI3Ba3E+wcB0XVOS8nJDqndg0=
github.com/mjl-/sconf v0.0.6 h1:5Dt58488ZOoVx680zgK2K3vUrokLsp5mXDUACrJlrUc= github.com/mjl-/sconf v0.0.6 h1:5Dt58488ZOoVx680zgK2K3vUrokLsp5mXDUACrJlrUc=
github.com/mjl-/sconf v0.0.6/go.mod h1:uF8OdWtLT8La3i4ln176i1pB0ps9pXGCaABEU55ZkE0= github.com/mjl-/sconf v0.0.6/go.mod h1:uF8OdWtLT8La3i4ln176i1pB0ps9pXGCaABEU55ZkE0=
github.com/mjl-/sherpa v0.6.7 h1:C5F8XQdV5nCuS4fvB+ye/ziUQrajEhOoj/t2w5T14BY= github.com/mjl-/sherpa v0.6.7 h1:C5F8XQdV5nCuS4fvB+ye/ziUQrajEhOoj/t2w5T14BY=

View file

@ -1,55 +1,12 @@
Bstore is a database library for storing and quering Go values. Bstore is an in-process database with serializable transactions supporting
referential/unique/nonzero constraints, (multikey) indices, automatic schema
management based on Go types and struct tags, and a query API.
Bstore is designed as a small, pure Go library that still provides most of Documentation:
the common data consistency requirements for modest database use cases. Bstore
aims to make basic use of cgo-based libraries, such as sqlite, unnecessary.
See https://pkg.go.dev/github.com/mjl-/bstore for features, examples and full https://pkg.go.dev/github.com/mjl-/bstore
documentation.
MIT-licensed MIT-licensed
Written by Mechiel Lukkien, mechiel@ueber.net. Feedback/bug reports/patches
# FAQ - Frequently Asked Questions welcome.
## Is bstore an ORM?
No. The API for bstore may look like an ORM. But instead of mapping bstore
"queries" (function calls) to an SQL query string, bstore executes them
directly without converting to a query language, storing the data itself.
## How does bstore store its data?
A bstore database is a single-file BoltDB database. BoltDB provides ACID
properties. Bstore uses a BoltDB "bucket" (key/value store) for each Go type
stored, with multiple subbuckets: one for type definitions, one for the actual
data, and one bucket per index. BoltDB stores data in a B+tree. See format.md
for details.
## How does bstore compare to sqlite?
Sqlite is a great library, but Go applications that require cgo are hard to
cross-compile. With bstore, cross-compiling to most Go-supported platforms
stays trivial (though not plan9, unfortunately). Although bstore is much more
limited in so many aspects than sqlite, bstore also offers some advantages as
well. Some points of comparison:
- Cross-compilation and reproducibility: Trivial with bstore due to pure Go,
much harder with sqlite because of cgo.
- Code complexity: low with bstore (7k lines including comments/docs), high
with sqlite.
- Query language: mostly-type-checked function calls in bstore, free-form query
strings only checked at runtime with sqlite.
- Functionality: very limited with bstore, much more full-featured with sqlite.
- Schema management: mostly automatic based on Go type definitions in bstore,
manual with ALTER statements in sqlite.
- Types and packing/parsing: automatic/transparent in bstore based on Go types
(including maps, slices, structs and custom MarshalBinary encoding), versus
manual scanning and parameter passing with sqlite with limited set of SQL
types.
- Performance: low to good performance with bstore, high performance with
sqlite.
- Database files: single file with bstore, several files with sqlite (due to
WAL or journal files).
- Test coverage: decent coverage but limited real-world for bstore, versus
extremely thoroughly tested and with enormous real-world use.

81
vendor/github.com/mjl-/bstore/doc.go generated vendored
View file

@ -1,16 +1,18 @@
/* /*
Package bstore is a database library for storing and querying Go values. Package bstore is an in-process database with serializable transactions
supporting referential/unique/nonzero constraints, (multikey) indices,
automatic schema management based on Go types and struct tags, and a query API.
Bstore is designed as a small, pure Go library that still provides most of Bstore a small, pure Go library that still provides most of the common data
the common data consistency requirements for modest database use cases. Bstore consistency requirements for modest database use cases. Bstore aims to make
aims to make basic use of cgo-based libraries, such as sqlite, unnecessary. basic use of cgo-based libraries, such as sqlite, unnecessary.
Bstore implements autoincrementing primary keys, indices, default values, Bstore implements autoincrementing primary keys, indices, default values,
enforcement of nonzero, unique and referential integrity constraints, automatic enforcement of nonzero, unique and referential integrity constraints, automatic
schema updates and a query API for combining filters/sorting/limits. Queries schema updates and a query API for combining filters/sorting/limits. Queries
are planned and executed using indices for fast execution where possible. are planned and executed using indices for speed where possible. Bstore works
Bstore is designed with the Go type system in mind: you typically don't have to with Go types: you typically don't have to write any (un)marshal code for your
write any (un)marshal code for your types. types. Bstore is not an ORM, it plans and executes queries itself.
# Field types # Field types
@ -142,37 +144,37 @@ Conversions that are not currently allowed, but may be in the future:
- Types of primary keys cannot be changed, also not from one integer type to a - Types of primary keys cannot be changed, also not from one integer type to a
wider integer type of same signedness. wider integer type of same signedness.
# BoltDB and storage # Bolt and storage
BoltDB is used as underlying storage. BoltDB stores key/values in a single Bolt is used as underlying storage through the bbolt library. Bolt stores
file, in multiple/nested buckets (namespaces) in a B+tree and provides ACID key/values in a single file, allowing multiple/nested buckets (namespaces) in a
transactions. Either a single write transaction or multiple read-only B+tree and provides ACID serializable transactions. A single write transaction
transactions can be active at a time. Do not start a blocking read-only can be active at a time, and one or more read-only transactions. Do not start
transaction while holding a writable transaction or vice versa, this will cause a blocking read-only transaction in a goroutine while holding a writable
deadlock. transaction or vice versa, this can cause deadlock.
BoltDB returns Go values that are memory mapped to the database file. This Bolt returns Go values that are memory mapped to the database file. This means
means BoltDB/bstore database files cannot be transferred between machines with Bolt/bstore database files cannot be transferred between machines with
different endianness. BoltDB uses explicit widths for its types, so files can different endianness. Bolt uses explicit widths for its types, so files can
be transferred between 32bit and 64bit machines of same endianness. While be transferred between 32bit and 64bit machines of same endianness. While
BoltDB returns read-only memory mapped byte slices, bstore only ever returns Bolt returns read-only memory mapped byte slices, bstore only ever returns
parsed/copied regular writable Go values that require no special programmer parsed/copied regular writable Go values that require no special programmer
attention. attention.
For each Go type opened for a database file, bstore ensures a BoltDB bucket For each Go type opened for a database file, bstore ensures a Bolt bucket
exists with two subbuckets: exists with two subbuckets:
- "types", with type descriptions of the stored records. Each time the database - "types", with type descriptions of the stored records. Each time the database
file is opened with a modified Go type (add/removed/modified file is opened with a modified Go type (add/removed/modified
field/type/bstore struct tag), a new type description is automatically added, field/type/bstore struct tag), a new type description is automatically added,
identified by sequence number. identified by sequence number.
- "records", containing all data, with the type's primary key as BoltDB key, - "records", containing all data, with the type's primary key as Bolt key,
and the encoded remaining fields as value. The encoding starts with a and the encoded remaining fields as value. The encoding starts with a
reference to a type description. reference to a type description.
For each index, another subbucket is created, its name starting with "index.". For each index, another subbucket is created, its name starting with "index.".
The stored keys consist of the index fields followed by the primary key, and an The stored keys consist of the index fields followed by the primary key, and an
empty value. empty value. See format.md for details.
# Limitations # Limitations
@ -189,12 +191,12 @@ equivalent of a nil pointer.
The first field of a stored struct is always the primary key. Autoincrement is The first field of a stored struct is always the primary key. Autoincrement is
only available for the primary key. only available for the primary key.
BoltDB opens the database file with a lock. Only one process can have the Bolt opens the database file with a lock. Only one process can have the
database open at a time. database open at a time.
An index stored on disk in BoltDB can consume more disk space than other An index stored on disk in Bolt can consume more disk space than other
database systems would: For each record, the indexed field(s) and primary key database systems would: For each record, the indexed field(s) and primary key
are stored in full. Because bstore uses BoltDB as key/value store, and doesn't are stored in full. Because bstore uses Bolt as key/value store, and doesn't
manage disk pages itself, it cannot as efficiently pack an index page with many manage disk pages itself, it cannot as efficiently pack an index page with many
records. records.
@ -202,5 +204,36 @@ Interface values cannot be stored. This would require storing the type along
with the value. Instead, use a type that is a BinaryMarshaler. with the value. Instead, use a type that is a BinaryMarshaler.
Values of builtin type "complex" cannot be stored. Values of builtin type "complex" cannot be stored.
Bstore inherits limitations from Bolt, see
https://pkg.go.dev/go.etcd.io/bbolt#readme-caveats-amp-limitations.
# Comparison with sqlite
Sqlite is a great library, but Go applications that require cgo are hard to
cross-compile. With bstore, cross-compiling to most Go-supported platforms
stays trivial (though not plan9, unfortunately). Although bstore is much more
limited in so many aspects than sqlite, bstore also offers some advantages as
well. Some points of comparison:
- Cross-compilation and reproducibility: Trivial with bstore due to pure Go,
much harder with sqlite because of cgo.
- Code complexity: low with bstore (7k lines including comments/docs), high
with sqlite.
- Query language: mostly-type-checked function calls in bstore, free-form query
strings only checked at runtime with sqlite.
- Functionality: very limited with bstore, much more full-featured with sqlite.
- Schema management: mostly automatic based on Go type definitions in bstore,
manual with ALTER statements in sqlite.
- Types and packing/parsing: automatic/transparent in bstore based on Go types
(including maps, slices, structs and custom MarshalBinary encoding), versus
manual scanning and parameter passing with sqlite with limited set of SQL
types.
- Performance: low to good performance with bstore, high performance with
sqlite.
- Database files: single file with bstore, several files with sqlite (due to
WAL or journal files).
- Test coverage: decent coverage but limited real-world for bstore, versus
extremely thoroughly tested and with enormous real-world use.
*/ */
package bstore package bstore

View file

@ -21,14 +21,31 @@ type exec[T any] struct {
// See plan.keys. We remove items from the list when we looked one up, but we keep the slice non-nil. // See plan.keys. We remove items from the list when we looked one up, but we keep the slice non-nil.
keys [][]byte keys [][]byte
// If -1, no limit is set. This is different from Query where 0 means // If non-empty, serve nextKey requests from here. Used when we need to do
// no limit. We count back and 0 means the end. // in-memory sort. After reading from here, and limit isn't reached yet, we may do
// another fill & sort of data to serve from, for orderings partially from an
// index. When filling data, limit (below) is accounted for, so all elements can be
// returned to caller.
data []pair[T]
// If -1, no limit is set. This is different from Query where 0 means no limit. We
// count back and 0 means the end. Also set from -1 to 0 when end of execution is
// reached.
limit int limit int
data []pair[T] // If not nil (even if empty), serve nextKey requests from here. // Index and record buckets loaded when first needed.
ib *bolt.Bucket ib *bolt.Bucket
rb *bolt.Bucket rb *bolt.Bucket
forward func() (bk, bv []byte) // Once we start scanning, we prepare forward to next/prev to the following value.
// Of last element in data. For finding end of group through prefix-match during
// partial index ordering for remaining in-memory sort.
lastik []byte
// If not nil, row that was scanned previously, to use instead of calling forward.
stowedbk, stowedbv []byte
// Once we start scanning, we prepare forward to next/prev to the following value.
forward func() (bk, bv []byte)
} }
// exec creates a new execution for the plan, registering statistics. // exec creates a new execution for the plan, registering statistics.
@ -54,7 +71,7 @@ func (p *plan[T]) exec(q *Query[T]) *exec[T] {
if len(p.orders) > 0 { if len(p.orders) > 0 {
q.stats.Sort++ q.stats.Sort++
} }
q.stats.LastOrdered = p.start != nil || p.stop != nil q.stats.LastOrdered = p.start != nil || p.stop != nil || p.norderidxuse > 0
q.stats.LastAsc = !p.desc q.stats.LastAsc = !p.desc
limit := -1 limit := -1
@ -107,12 +124,9 @@ func (e *exec[T]) nextKey(write, value bool) ([]byte, T, error) {
return nil, zero, q.err return nil, zero, q.err
} }
// We collected & sorted data previously. Return from it until done. // We collected & sorted data previously.
// Limit was already applied. // Limit was already applied/updated, so we can serve these without checking.
if e.data != nil { if len(e.data) > 0 {
if len(e.data) == 0 {
return nil, zero, ErrAbsent
}
p := e.data[0] p := e.data[0]
e.data = e.data[1:] e.data = e.data[1:]
var v T var v T
@ -127,6 +141,7 @@ func (e *exec[T]) nextKey(write, value bool) ([]byte, T, error) {
return p.bk, v, nil return p.bk, v, nil
} }
// Limit is 0 when we hit the limit or at end of processing the execution.
if e.limit == 0 { if e.limit == 0 {
return nil, zero, ErrAbsent return nil, zero, ErrAbsent
} }
@ -153,10 +168,8 @@ func (e *exec[T]) nextKey(write, value bool) ([]byte, T, error) {
// List of IDs (records) or full unique index equality match. // List of IDs (records) or full unique index equality match.
// We can get the records/index value by a simple "get" on the key. // We can get the records/index value by a simple "get" on the key.
if e.keys != nil { if e.keys != nil {
// If we need to sort, we collect all elements and prevent further querying.
collect := len(e.plan.orders) > 0 collect := len(e.plan.orders) > 0
if collect {
e.data = []pair[T]{} // Must be non-nil to get into e.data branch!
}
for i, xk := range e.keys { for i, xk := range e.keys {
var bk, bv []byte var bk, bv []byte
@ -217,6 +230,7 @@ func (e *exec[T]) nextKey(write, value bool) ([]byte, T, error) {
return bk, v, nil return bk, v, nil
} }
if !collect { if !collect {
e.limit = 0
return nil, zero, ErrAbsent return nil, zero, ErrAbsent
} }
// Restart, now with data. // Restart, now with data.
@ -225,14 +239,13 @@ func (e *exec[T]) nextKey(write, value bool) ([]byte, T, error) {
if e.limit > 0 && len(e.data) > e.limit { if e.limit > 0 && len(e.data) > e.limit {
e.data = e.data[:e.limit] e.data = e.data[:e.limit]
} }
e.limit = 0
return q.nextKey(write, value) return q.nextKey(write, value)
} }
// We are going to do a scan, either over the records or an index. We may have a start and stop key. // We are going to do a scan, either over the records or (a part of) an index. We
// may have a start and stop key.
collect := len(e.plan.orders) > 0 collect := len(e.plan.orders) > 0
if collect {
e.data = []pair[T]{} // Must be non-nil to get into e.data branch on function restart.
}
// Every 1k keys we've seen, we'll check if the context has been canceled. If we // Every 1k keys we've seen, we'll check if the context has been canceled. If we
// wouldn't do this, a query that doesn't return any matches won't get canceled // wouldn't do this, a query that doesn't return any matches won't get canceled
// until it is finished. // until it is finished.
@ -304,6 +317,10 @@ func (e *exec[T]) nextKey(write, value bool) ([]byte, T, error) {
} }
} }
} }
} else if e.stowedbk != nil {
// Resume with previously seen key/value.
xk, xv = e.stowedbk, e.stowedbv
e.stowedbk, e.stowedbv = nil, nil
} else { } else {
if e.plan.idx == nil { if e.plan.idx == nil {
q.stats.Records.Cursor++ q.stats.Records.Cursor++
@ -331,24 +348,37 @@ func (e *exec[T]) nextKey(write, value bool) ([]byte, T, error) {
} }
var pk, bv []byte var pk, bv []byte
ordersidxPartial := e.plan.norderidxuse > 0 && len(e.plan.orders) > 0
var idxkeys [][]byte // Only set when we have partial ordering from index.
if e.plan.idx == nil { if e.plan.idx == nil {
pk = xk pk = xk
bv = xv bv = xv
} else { } else {
var err error var err error
pk, _, err = e.plan.idx.parseKey(xk, false) pk, idxkeys, err = e.plan.idx.parseKey(xk, ordersidxPartial, true)
if err != nil { if err != nil {
q.error(err) q.error(err)
return nil, zero, err return nil, zero, err
} }
} }
// If we have a parial order from the index, and this new value has a different
// index ordering key prefix than the last value, we stop collecting, sort the data we
// have by the remaining ordering, return that data, and continue collecting in the
// next round. We stow the new value so we don't have to revert the forward() from
// earlier.
if ordersidxPartial && len(e.data) > 0 && !prefixMatch(e.lastik, e.plan.norderidxuse, idxkeys, pk) {
e.stowedbk, e.stowedbv = xk, xv
break
}
p := pair[T]{pk, bv, nil} p := pair[T]{pk, bv, nil}
if ok, err := e.checkFilter(&p); err != nil { if ok, err := e.checkFilter(&p); err != nil {
return nil, zero, err return nil, zero, err
} else if !ok { } else if !ok {
continue continue
} }
//log.Printf("have kv, %x %x", p.bk, p.bv) //log.Printf("have kv, %x %x", p.bk, p.bv)
var v T var v T
var err error var err error
@ -361,6 +391,7 @@ func (e *exec[T]) nextKey(write, value bool) ([]byte, T, error) {
} }
if collect { if collect {
e.data = append(e.data, p) e.data = append(e.data, p)
e.lastik = xk
continue continue
} }
if e.limit > 0 { if e.limit > 0 {
@ -368,17 +399,39 @@ func (e *exec[T]) nextKey(write, value bool) ([]byte, T, error) {
} }
return p.bk, v, nil return p.bk, v, nil
} }
if !collect { if !collect || len(e.data) == 0 {
e.limit = 0
return nil, zero, ErrAbsent return nil, zero, ErrAbsent
} }
// Restart, now with data. // Restart, now with data.
e.sort() e.sort()
if e.limit > 0 && len(e.data) > e.limit { if e.limit > 0 {
if len(e.data) > e.limit {
e.data = e.data[:e.limit] e.data = e.data[:e.limit]
} }
e.limit -= len(e.data)
}
return e.nextKey(write, value) return e.nextKey(write, value)
} }
// prefixMatch returns whether ik (index key) starts with the bytes from n elements
// from field keys kl and primary key pk.
func prefixMatch(ik []byte, n int, kl [][]byte, pk []byte) bool {
for i := 0; i < n; i++ {
var k []byte
if i < len(kl) {
k = kl[i]
} else {
k = pk
}
if !bytes.HasPrefix(ik, k) {
return false
}
ik = ik[len(k):]
}
return true
}
// checkFilter checks against the filters for the plan. // checkFilter checks against the filters for the plan.
func (e *exec[T]) checkFilter(p *pair[T]) (rok bool, rerr error) { func (e *exec[T]) checkFilter(p *pair[T]) (rok bool, rerr error) {
q := e.q q := e.q
@ -574,6 +627,10 @@ func compare(k kind, a, b reflect.Value) int {
} }
func (e *exec[T]) sort() { func (e *exec[T]) sort() {
if len(e.data) <= 1 {
return
}
// todo: We should check whether we actually need to load values. We're // todo: We should check whether we actually need to load values. We're
// always loading it for the time being because SortStableFunc isn't // always loading it for the time being because SortStableFunc isn't
// going to give us a *pair (even though it could because of the slice) // going to give us a *pair (even though it could because of the slice)

View file

@ -329,8 +329,7 @@ func (ft fieldType) parseValue(p *parser) any {
if fm.Nonzero(i) { if fm.Nonzero(i) {
l = append(l, ft.ListElem.parseValue(p)) l = append(l, ft.ListElem.parseValue(p))
} else { } else {
// Always add non-zero elements, or we would // Always add zero elements, or we would change the number of elements in a list.
// change the number of elements in a list.
l = append(l, ft.ListElem.zeroExportValue()) l = append(l, ft.ListElem.zeroExportValue())
} }
} }
@ -343,8 +342,8 @@ func (ft fieldType) parseValue(p *parser) any {
if fm.Nonzero(i) { if fm.Nonzero(i) {
l[i] = ft.ListElem.parseValue(p) l[i] = ft.ListElem.parseValue(p)
} else { } else {
// Always add non-zero elements, or we would // Always add zero elements, or we would change the number of elements in the
// change the number of elements in a list. // array.
l[i] = ft.ListElem.zeroExportValue() l[i] = ft.ListElem.zeroExportValue()
} }
} }

View file

@ -132,7 +132,8 @@ func parsePK(rv reflect.Value, bk []byte) error {
// parseKey parses the PK (last element) of an index key. // parseKey parses the PK (last element) of an index key.
// If all is set, also gathers the values before and returns them in the second // If all is set, also gathers the values before and returns them in the second
// parameter. // parameter.
func (idx *index) parseKey(buf []byte, all bool) ([]byte, [][]byte, error) { // If witnull is set, string values will get their ending \0 included.
func (idx *index) parseKey(buf []byte, all bool, withnull bool) ([]byte, [][]byte, error) {
var err error var err error
var keys [][]byte var keys [][]byte
take := func(n int) { take := func(n int) {
@ -160,7 +161,11 @@ fields:
for i, b := range buf { for i, b := range buf {
if b == 0 { if b == 0 {
if all { if all {
keys = append(keys, buf[:i]) o := i
if withnull {
o++
}
keys = append(keys, buf[:o])
} }
buf = buf[i+1:] buf = buf[i+1:]
continue fields continue fields

View file

@ -8,7 +8,6 @@ import (
) )
// todo: cache query plans? perhaps explicitly through something like a prepared statement. the current plan includes values in keys,start,stop, which would need to be calculated for each execution. should benchmark time spent in planning first. // todo: cache query plans? perhaps explicitly through something like a prepared statement. the current plan includes values in keys,start,stop, which would need to be calculated for each execution. should benchmark time spent in planning first.
// todo optimize: handle multiple sorts with multikey indices if they match
// todo optimize: combine multiple filter (not)in/equals calls for same field // todo optimize: combine multiple filter (not)in/equals calls for same field
// todo optimize: efficiently pack booleans in an index (eg for Message.Flags), and use it to query. // todo optimize: efficiently pack booleans in an index (eg for Message.Flags), and use it to query.
// todo optimize: do multiple range scans if necessary when we can use an index for an equal check with multiple values. // todo optimize: do multiple range scans if necessary when we can use an index for an equal check with multiple values.
@ -32,9 +31,14 @@ type plan[T any] struct {
keys [][]byte keys [][]byte
desc bool // Direction of the range scan. desc bool // Direction of the range scan.
start []byte // First key to scan. Filters below may still apply. If desc, this value is > than stop (if it is set). If nil, we begin ranging at the first or last (for desc) key. // First key to scan. Filters below may still apply. If desc, this value is > than
stop []byte // Last key to scan. Can be nil independently of start. // stop (if it is set). If nil, we begin ranging at the first or last (for desc)
startInclusive bool // If the start and stop values are inclusive or exclusive. // key.
start []byte
// Last key to scan. Can be nil independently of start.
stop []byte
// If the start and stop values are inclusive or exclusive.
startInclusive bool
stopInclusive bool stopInclusive bool
// Filter we need to apply after retrieving the record. If all original filters // Filter we need to apply after retrieving the record. If all original filters
@ -42,9 +46,13 @@ type plan[T any] struct {
// empty. // empty.
filters []filter[T] filters []filter[T]
// Orders we need to apply after first retrieving all records. As with // Number of fields from index used to group results before applying in-memory
// filters, if a range scan takes care of an ordering from the query, // ordering with "orders" below.
// this field is empty. norderidxuse int
// Orders we need to apply after first retrieving all records with equal values for
// first norderidxuse fields. As with filters, if a range scan takes care of all
// orderings from the query, this field is empty.
orders []order orders []order
} }
@ -177,7 +185,7 @@ indices:
var p *plan[T] var p *plan[T]
var nexact int var nexact int
var nrange int var nrange int
var ordered bool var norder int
evaluatePKOrIndex := func(idx *index) error { evaluatePKOrIndex := func(idx *index) error {
var isPK bool var isPK bool
@ -205,25 +213,40 @@ indices:
} }
var nex = 0 var nex = 0
// log.Printf("idx %v", idx) // log.Printf("evaluating idx %#v", idx)
var skipFilters []*filter[T] var skipFilters []*filter[T]
for _, f := range idx.Fields { for _, f := range idx.Fields {
if equals[f.Name] != nil && f.Type.Kind != kindSlice { if equals[f.Name] != nil && f.Type.Kind != kindSlice {
skipFilters = append(skipFilters, equals[f.Name]) skipFilters = append(skipFilters, equals[f.Name])
nex++
} else if inslices[f.Name] != nil && f.Type.Kind == kindSlice { } else if inslices[f.Name] != nil && f.Type.Kind == kindSlice {
skipFilters = append(skipFilters, inslices[f.Name]) skipFilters = append(skipFilters, inslices[f.Name])
nex++
} else { } else {
break break
} }
nex++
} }
// See if the next field can be used for compare. // For ordering, skip leading filters we already match on exactly.
var gx, lx *filterCompare[T]
var nrng int
var order *order
orders := q.xorders orders := q.xorders
trim := 0
TrimOrders:
for _, o := range orders {
for _, f := range idx.Fields[:nex] {
if o.field.Name == f.Name {
trim++
continue TrimOrders
}
}
break
}
orders = orders[trim:]
// Fields from the index that we use for grouping before in-memory sorting.
var norderidxuse int
// See if the next index field can be used for compare and ordering.
var gx, lx *filterCompare[T]
var nrng int // for nrange
if nex < len(idx.Fields) { if nex < len(idx.Fields) {
nf := idx.Fields[nex] nf := idx.Fields[nex]
for i := range q.xfilters { for i := range q.xfilters {
@ -250,22 +273,45 @@ indices:
} }
} }
// See if it can be used for ordering. // We can use multiple orderings as long as the asc/desc direction stays the same.
// todo optimize: we could use multiple orders nord := 0
if len(orders) > 0 && orders[0].field.Name == nf.Name { for i, o := range orders {
order = &orders[0] if nex+i < len(idx.Fields) && o.field.Name == idx.Fields[nex+i].Name && (nord == 0 || o.asc == orders[0].asc) {
orders = orders[1:] nord++
continue
} }
break
}
norderidxuse = nex + nord
prevorders := orders
orders = orders[nord:]
// The stored index key ends with the primary key, so if we're there, and the next
// ordering key is the primary key, we use the index for it too.
if norderidxuse == len(idx.Fields) && len(orders) > 0 && orders[0].field.Name == q.st.Current.Fields[0].Name && (nord == 0 || orders[0].asc == prevorders[nord-1].asc) {
orders = orders[1:]
norderidxuse++
}
} else if len(orders) > 0 && orders[0].field.Name == q.st.Current.Fields[0].Name {
// We only had equals filters that used all of the index, but we're also sorting by
// the primary key, so use the index for that too.
orders = orders[1:]
norderidxuse++
} }
// Orders handled by the index, excluding exact match filters.
idxorders := q.xorders[trim : len(q.xorders)-len(orders)]
// log.Printf("index fields to match for index order: %d, orders for index %d, in-memory ordering %d, total orders %d", norderidxuse, len(idxorders), len(orders), len(q.xorders))
// See if this is better than what we had. // See if this is better than what we had.
if !(nex > nexact || (nex == nexact && (nrng > nrange || order != nil && !ordered && (q.xlimit > 0 || nrng == nrange)))) { if !(nex > nexact || (nex == nexact && (nrng > nrange || len(idxorders) > norder && (q.xlimit > 0 || nrng == nrange)))) {
// log.Printf("plan not better, nex %d, nrng %d, limit %d, order %v ordered %v", nex, nrng, q.limit, order, ordered) // log.Printf("plan not better, nex %d, nrng %d, limit %d, nidxorders %v ordered %v", nex, nrng, q.xlimit, len(idxorders), norder)
return nil return nil
} }
nexact = nex nexact = nex
nrange = nrng nrange = nrng
ordered = order != nil norder = len(idxorders)
// Calculate the prefix key. // Calculate the prefix key.
var kvalues []reflect.Value var kvalues []reflect.Value
@ -307,7 +353,8 @@ indices:
startInclusive := gx == nil || gx.op != opGreater startInclusive := gx == nil || gx.op != opGreater
stopInclusive := lx == nil || lx.op != opLess stopInclusive := lx == nil || lx.op != opLess
if order != nil && !order.asc { desc := len(idxorders) > 0 && !idxorders[0].asc
if desc {
start, stop = stop, start start, stop = stop, start
startInclusive, stopInclusive = stopInclusive, startInclusive startInclusive, stopInclusive = stopInclusive, startInclusive
} }
@ -318,12 +365,13 @@ indices:
p = &plan[T]{ p = &plan[T]{
idx: idx, idx: idx,
desc: order != nil && !order.asc, desc: desc,
start: start, start: start,
stop: stop, stop: stop,
startInclusive: startInclusive, startInclusive: startInclusive,
stopInclusive: stopInclusive, stopInclusive: stopInclusive,
filters: dropFilters(q.xfilters, skipFilters), filters: dropFilters(q.xfilters, skipFilters),
norderidxuse: norderidxuse,
orders: orders, orders: orders,
} }
return nil return nil
@ -341,6 +389,7 @@ indices:
} }
if p != nil { if p != nil {
// log.Printf("using index plan %v", p)
return p, nil return p, nil
} }

View file

@ -561,7 +561,7 @@ func (db *DB) Register(ctx context.Context, typeValues ...any) error {
parsePK(a, prev.buf[prev.pre:]) // Ignore error, nothing to do. parsePK(a, prev.buf[prev.pre:]) // Ignore error, nothing to do.
parsePK(b, k.buf[k.pre:]) // Ignore error, nothing to do. parsePK(b, k.buf[k.pre:]) // Ignore error, nothing to do.
var dup []any var dup []any
_, values, _ := idx.parseKey(k.buf, true) _, values, _ := idx.parseKey(k.buf, true, false)
for i := range values { for i := range values {
x := reflect.New(reflect.TypeOf(idx.Fields[i].Type.zeroKey())).Elem() x := reflect.New(reflect.TypeOf(idx.Fields[i].Type.zeroKey())).Elem()
parsePK(x, values[i]) // Ignore error, nothing to do. parsePK(x, values[i]) // Ignore error, nothing to do.

2
vendor/modules.txt vendored
View file

@ -16,7 +16,7 @@ github.com/mjl-/adns/internal/singleflight
# github.com/mjl-/autocert v0.0.0-20231214125928-31b7400acb05 # github.com/mjl-/autocert v0.0.0-20231214125928-31b7400acb05
## explicit; go 1.20 ## explicit; go 1.20
github.com/mjl-/autocert github.com/mjl-/autocert
# github.com/mjl-/bstore v0.0.4 # github.com/mjl-/bstore v0.0.5
## explicit; go 1.19 ## explicit; go 1.19
github.com/mjl-/bstore github.com/mjl-/bstore
# github.com/mjl-/sconf v0.0.6 # github.com/mjl-/sconf v0.0.6