disktable.go

/*
Package disktable provides a write-once, read-many table with index supoprt.
This is build on top of badgerDB, which is basically a key/value SSTable storage mechanism.

Let's create a table with some data:

	dir := filepath.Join(os.TempDir(), "your_table"")
	// Remove it if exists, may or may not want to do this. However you cannot
	// create a table on a directory that exists.
	os.RemoveAll(dir)

	// These are our indexes on the data. AllowDuplicates allows duplicate entries
	// in the index.
	indexes := NewIndexes(
		&Index{Name: "First Name", AllowDuplicates: true},
		&Index{Name: "Last Name", AllowDuplicates: true},
		&Index{Name: "ID"},
	}

	w, err := New(dir, WithIndexes(indexes))
	if err != nil {
		panic(err)
	}

	for _, data := range someData {
		b, err := proto.Marshal(data)
		if err != nil {
			panic(err)
		}

		insert := indexes.Insert(b).AddIndexKey(
			"First Name", UnsafeGetBytes(data.First),
		).AddIndexKey(
			"Last Name", UnsafeGetBytes(data.Last),
		).AddIndexKey(
			"ID", NumToByte(data.ID),
		)

		if err = w.WriteData(insert); err != nil {
			panic(err)
		}
	}

	if err := w.Close(); err != nil {
		panic(err)
	}

Now let's open it and stream all records:

	table, err := Open(dir)
	if err != nil {
		panic(err)
	}

	results, err := table.FetchAll(ctx)
	if err != nil {
		panic(err)
	}

	for result := range results {
		if result.Err != nil {
			panic(err)
		}

		entry := &pb.MyData{}
		if err := proto.Unmarshal(entry, result.Value); err != nil {
			panic(err)
		}

		fmt.Println("found: ", pretty.Sprint(entry))
	}

Let's look for all entries that have the first name John:

	results, err := table.Fetch(
		ctx,
		Lookup{IndexName: "First Name", Key: UnsafeGetBytes("John")},
	)

	if err != nil {
		panic(err)
	}

	for result := range results {
		if result.Err != nil {
			panic(err)
		}

		entry := &pb.MyData{}
		if err := proto.Unmarshal(entry, result.Value); err != nil {
			panic(err)
		}

		fmt.Println("found: ", pretty.Sprint(entry))
	}
*/
package disktable

import (
	"bytes"
	"context"
	"encoding/binary"
	"fmt"
	"math"
	"os"
	"path/filepath"
	"reflect"
	"runtime"
	"sync"
	"sync/atomic"
	"unsafe"

	// This marks the outcaste-io version, which is newer but gets rid of the values file.
	// badger is stable, so at this point I see no reason to move.
	/*
		badger "github.com/outcaste-io/badger/v3"
		"github.com/outcaste-io/ristretto/z"
	*/

	badger "github.com/dgraph-io/badger/v3"
	"github.com/dgraph-io/ristretto/z"
	"github.com/google/btree"
	"github.com/johnsiilver/calloptions"
	"golang.org/x/sync/errgroup"
)

var endian = binary.BigEndian

// Table represents our read-only table.
type Table struct {
	// primary is the primary data table.
	primary     *Index
	indexByName map[string]*Index
	// indexes is the indexes in the order they were added.
	indexes []*Index

	logger badger.Logger
}

// OpenOption is optional arguments for Open().
type OpenOption interface {
	openOption()
}

var readOpts = badger.DefaultOptions(
	"",
).WithBlockCacheSize(
	2147483648,
).WithNumGoroutines(
	runtime.NumCPU(),
).WithReadOnly(
	true,
).WithBloomFalsePositive(
	.01,
)

// Open opens an existing disktable for reading.
func Open(pathDir string, options ...OpenOption) (*Table, error) {
	if _, err := os.Stat(pathDir); err != nil {
		return nil, fmt.Errorf("pathDir %q does not exist", pathDir)
	}

	table := &Table{indexByName: map[string]*Index{}, logger: nullLogger{}}

	if err := calloptions.ApplyOptions(table, options); err != nil {
		return nil, err
	}

	td := tableDef{}
	if err := td.unmarshal(pathDir); err != nil {
		return nil, err
	}
	counter := atomic.Uint64{}
	counter.Store(td.Length)

	// Open all our index tables.
	for _, col := range td.Columns {
		p := filepath.Join(pathDir, col.Path)

		badgeOpts := readOpts.WithDir(p).WithValueDir(p).WithLogger(table.logger)
		db, err := badger.Open(badgeOpts)
		if err != nil {
			return nil, err
		}

		index := &Index{
			Name:            col.Name,
			AllowDuplicates: col.AllowDuplicates,
			db:              db,
			buffPool:        make(chan *z.Buffer, 100),
			counter:         counter,
		}
		table.indexes = append(table.indexes, index)
		table.indexByName[col.Name] = index
	}

	// Open our primary data table.
	primaryPath := filepath.Join(pathDir, "primary")
	badgeOpts := readOpts.WithDir(primaryPath).WithValueDir(primaryPath).WithLogger(table.logger)
	db, err := badger.Open(badgeOpts)
	if err != nil {
		return nil, err
	}
	table.primary = &Index{
		Name:     "primary",
		db:       db,
		buffPool: make(chan *z.Buffer, 100),
		counter:  counter,
	}

	return table, nil
}

// Close closes all the databases.
func (t *Table) Close() error {
	err := t.primary.db.Close()

	for _, index := range t.indexes {
		if e := index.db.Close(); e != nil && err == nil {
			err = e
		}
	}
	return err
}

// Get gets the i'th entry stored in the table.
func (t *Table) Get(ctx context.Context, i uint64) ([]byte, error) {
	// The primary index is stored with the starting index as 1. Externally
	// we use the standard starting index of 0. So we i++.
	i++

	var v []byte
	err := t.primary.db.View(
		func(txn *badger.Txn) error {
			item, err := txn.Get(NumToByte(i))
			if err != nil {
				return err
			}

			v, err = item.ValueCopy(nil)
			if err != nil {
				return fmt.Errorf("problem getting value at index %d: %w", i, err)
			}
			return nil
		},
	)
	if err != nil {
		return nil, err
	}
	return v, nil
}

// Len() returns the number of entries in the table.
func (t *Table) Len() uint64 {
	return t.primary.counter.Load()
}

// Lookup provides the Index name and the Value that needs to match for the entry
// to be returned.
type Lookup struct {
	// IndexName is the name of the index to do the lookup in.
	IndexName string
	// Key is the key in the index to lookup.
	Key []byte
}

// Result is the result of a table lookup.
type Result struct {
	Value []byte
	Err   error
}

type btreeItem struct {
	primaryKey uint64
	count      uint16
}

func (a btreeItem) Less(than btree.Item) bool {
	b := than.(btreeItem)
	return a.primaryKey < b.primaryKey
}

/*
Fetch retrieves specifc rows that match all index lookups. You cannot currently
specify multiple searches in the same index. If you wish to fetch all rows, use FetchAll().
Here is an example:

	results, err := table.Fetch(
		ctx,
		Lookup{IndexName: "First Name", Key: UnsafeGetBytes("John")},
	)

	if err != nil {
		panic(err)
	}

	for result := range results {
		if result.Err != nil {
			panic(err)
		}

		entry := &pb.MyData{}
		if err := proto.Unmarshal(entry, result.Value); err != nil {
			panic(err)
		}

		fmt.Println("found: ", pretty.Sprint(entry))
	}
*/
func (t *Table) Fetch(ctx context.Context, primary Lookup, secondaries ...Lookup) (chan Result, error) {
	indexesSearched := map[string]bool{}
	lookups := append(secondaries, primary)

	for _, l := range lookups {
		if t.indexByName[l.IndexName] == nil {
			return nil, fmt.Errorf("index %q is not found", l.IndexName)
		}
		if indexesSearched[l.IndexName] {
			return nil, fmt.Errorf("index %q cannot be used in a search more than once", l.IndexName)
		}
		if len(l.Key) == 0 {
			return nil, fmt.Errorf("index %q lookup cannot have an empty key", l.IndexName)
		}
		indexesSearched[l.IndexName] = true
	}

	ch := make(chan Result, 1)

	g, ctx := errgroup.WithContext(ctx)
	mu := &sync.Mutex{}
	matches := btree.New(2)

	/*
		TODO(jdoak): This could be made faster. Instead we could do:
		- Look for an index with non-duplicates (if non, default to the primary Lookup)
		- For ever match in that index, check the other indexes using a goroutine pool
		- If they match, write to batch block (which is now a btree)
		- If len(batch block) == 100, do a fetch and return in another goroutine
	*/
	go func() {
		defer close(ch)
		for _, l := range lookups {
			l := l
			g.Go(
				func() error {
					return t.insertResultsFromIndex(ctx, mu, l, matches)
				},
			)
		}
		if err := g.Wait(); err != nil {
			ch <- Result{Err: err}
			return
		}

		keysLookup := make([][]byte, 0, 100)

		// Pull all the rows in the index that matched and send them back.
		matches.Ascend(
			func(item btree.Item) bool {
				i := item.(btreeItem)
				if int(i.count) == len(lookups) {
					primaryKey := make([]byte, 8)
					endian.PutUint64(primaryKey, i.primaryKey)
					keysLookup = append(keysLookup, primaryKey)
					// Ok, we have enough for a bulk lookup.
					if len(keysLookup) == 1000 {
						t.primary.db.View(t.retrievePrimaryRows(keysLookup, ch))
						keysLookup = keysLookup[0:0]
					}
				}
				return true
			},
		)

		// Get any outstanding lookups that didn't have enough to do a bulk.
		if len(keysLookup) > 0 {
			t.primary.db.View(t.retrievePrimaryRows(keysLookup, ch))
		}
	}()

	return ch, nil
}

func (t *Table) insertResultsFromIndex(ctx context.Context, mu *sync.Mutex, lookup Lookup, matches *btree.BTree) error {
	for result := range t.indexLookup(lookup.IndexName, lookup.Key) {
		if ctx.Err() != nil {
			return ctx.Err()
		}
		if result.err != nil {
			if result.err == badger.ErrKeyNotFound {
				continue
			}
			return result.err
		}
		// Add entry to btree or increase existing entry counter.
		mu.Lock()

		var counter btreeItem
		a := matches.Get(btreeItem{primaryKey: result.i})
		if a == nil {
			counter = btreeItem{primaryKey: result.i}
		} else {
			counter = a.(btreeItem)
		}
		counter.count++
		matches.ReplaceOrInsert(counter)

		mu.Unlock()
	}
	return nil
}

func (t *Table) retrievePrimaryRows(keys [][]byte, ch chan Result) func(txn *badger.Txn) error {
	return func(txn *badger.Txn) error {
		for _, key := range keys {
			item, err := txn.Get(key)
			if err != nil {
				return err
			}
			if !bytes.Equal(item.Key(), key) {
				panic("known key search came up with different key: HUGE PROBLEM, HUGE!!!!!!")
			}
			value, err := item.ValueCopy(nil)
			if err != nil {
				return fmt.Errorf("problem getting value of key(%s): %w", string(key), err)
			}
			ch <- Result{Value: value}
		}
		return nil
	}
}

type lookupResult struct {
	i   uint64
	err error
}

// indexLookup returns all matching keys in an index.
func (t *Table) indexLookup(indexName string, key []byte) chan lookupResult {
	index := t.indexByName[indexName]

	realKey := finalKey(key, 0)

	ch := make(chan lookupResult, 1)
	if index.AllowDuplicates {
		// multiMatch will close the channel.
		go t.multiMatch(realKey, index, ch)
	} else {
		ch <- t.exactMatch(realKey, index)
		close(ch)
	}

	return ch
}

// exactMatch does an exact match for a key in the index. key must have been finalKey()'d.
func (t *Table) exactMatch(key []byte, index *Index) lookupResult {
	var result lookupResult

	f := func(txn *badger.Txn) error {
		item, err := txn.Get(key)
		if err != nil {
			return err
		}
		if !bytes.Equal(item.Key(), key) {
			panic("known key search came up with different key: HUGE PROBLEM, HUGE!!!!!!")
		}
		value, err := item.ValueCopy(nil)
		if err != nil {
			return fmt.Errorf("problem getting value of key(%s): %w", string(key), err)
		}
		result = lookupResult{i: endian.Uint64(value)}
		return nil
	}

	if err := index.db.View(f); err != nil {
		result.err = err
	}
	return result

}

// multiMatch matches all keys in an index that allows duplicates. key must have been finalKey()'d.
func (t *Table) multiMatch(key []byte, index *Index, ch chan lookupResult) {
	defer close(ch)

	keyData, err := getPrefixSuffix(key)
	if err != nil {
		ch <- lookupResult{err: err}
		return
	}

	tempBuff := make([]byte, 8)
	err = index.db.View(
		func(txn *badger.Txn) error {
			it := txn.NewIterator(badger.DefaultIteratorOptions)
			defer it.Close()

			for it.Seek(keyData.prefix); it.ValidForPrefix(keyData.prefix); it.Next() {
				item := it.Item()

				foundKeyData, err := getPrefixSuffix(item.Key())
				if err != nil {
					return fmt.Errorf("multiMatch of index(%s) had bad key(%v): %w", index.Name, item.Key(), err)
				}
				// It is possible to have a prefix match for something not exactly the same.
				// But we are looking for exact prefix matches followed by a number followed by
				// our key size. By matching prefix and key size, we can be sure of an exact match.
				// If not, throw out the data.
				if len(foundKeyData.prefix) != len(keyData.prefix) {
					continue
				}
				b, err := item.ValueCopy(tempBuff)
				if err != nil {
					return err
				}
				ch <- lookupResult{i: endian.Uint64(b)}
			}
			return nil
		},
	)
	if err != nil {
		ch <- lookupResult{err: err}
	}
}

type FetchAllOption interface {
	fetchAll()
}

type fetchAllOptions struct {
	numGo int
}

func (f *fetchAllOptions) defaults() {
	if f.numGo == 0 {
		f.numGo = 16
	}
}

// NumStreamGoroutines sets the number of goroutines to be used in FetchAll(). By
// default this is 16.
func NumStreamGoroutines(n int) interface {
	FetchAllOption
	calloptions.CallOption
} {
	return struct {
		FetchAllOption
		calloptions.CallOption
	}{
		CallOption: calloptions.New(
			func(a any) error {
				x := a.(*fetchAllOptions)
				x.numGo = n
				return nil
			},
		),
	}
}

// FetchAll fetches all the tables entries.
func (t *Table) FetchAll(ctx context.Context, options ...FetchAllOption) (chan Result, error) {
	opts := fetchAllOptions{}
	if err := calloptions.ApplyOptions(opts, options); err != nil {
		return nil, err
	}
	opts.defaults()

	stream := t.primary.db.NewStream()
	stream.NumGo = opts.numGo
	stream.LogPrefix = "FetchAll.Streaming"

	ch := make(chan Result, 1)
	stream.Send = func(buf *z.Buffer) error {
		list, err := badger.BufferToKVList(buf)
		if err != nil {
			return err
		}

		for _, kv := range list.Kv {
			// TODO(jdoak): This doesn't seem to matter.
			v := make([]byte, len(kv.Value))
			copy(v, kv.Value)
			select {
			case <-ctx.Done():
				return ctx.Err()
			case ch <- Result{Value: v}:
			}
		}
		return nil
	}

	go func() {
		defer close(ch)
		if err := stream.Orchestrate(context.Background()); err != nil {
			ch <- Result{Err: err}
			return
		}
	}()

	return ch, nil
}

// ByteSlice2String coverts a []byte to a string without incurring the cost of a copy of
// the given []byte parameter. This is an unsafe operation and requires that you never
// modify the []byte slice you passed in.
func ByteSlice2String(bs []byte) string {
	if len(bs) == 0 {
		return ""
	}
	return *(*string)(unsafe.Pointer(&bs))
}

// UnsafeGetBytes retrieves the underlying []byte held in string "s" without doing
// a copy. Do not modify the []byte or suffer the consequences.
func UnsafeGetBytes(s string) []byte {
	if s == "" {
		return nil
	}
	return (*[0x7fff0000]byte)(unsafe.Pointer(
		(*reflect.StringHeader)(unsafe.Pointer(&s)).Data),
	)[:len(s):len(s)]
}

// Number represents any uint*, int* or float* type.
type Number interface {
	~uint | ~uint8 | ~uint16 | ~uint32 | ~uint64 |
		~int | ~int8 | ~int16 | ~int32 | ~int64 |
		~float32 | ~float64
}

// NumToByte converts a number into a BigEndian []byte sequence.
func NumToByte[N Number](n N) []byte {
	buff := bytes.Buffer{}

	binary.Write(&buff, binary.BigEndian, n)
	return buff.Bytes()
}

// ByteToNum returns a number stored in b that represents N. That number should
// be encoded in BigEndian, usually by NumToByte().
func ByteToNum[N Number](b []byte) (N, error) {
	var v N
	if err := binary.Read(bytes.NewReader(b), binary.BigEndian, v); err != nil {
		return 0, err
	}
	return v, nil
}

// finalKey creates a final key out of a key and a counter. If len(key) == 0 and counter > 0,
// it returns counter as a [8]byte.
func finalKey(key []byte, counter uint64) []byte {
	var x []byte
	switch {
	case counter == 0 && len(key) == 0:
		panic("bug: can't have a finalKey() call with key == nil  and counter == 0")
	case counter == 0:
		// [key][key size(uint16)]
		x = make([]byte, 2+len(key))
		copy(x, key)
		endian.PutUint16(x[len(key):], uint16(len(key)))
		return x
	case len(key) == 0:
		// [counter(uint64)]
		x = make([]byte, 8)
		endian.PutUint64(x, counter)
		return x
	}
	// [key][counter(uint64)][key size(uint16)]
	x = make([]byte, 2+8+len(key))
	copy(x, key)
	endian.PutUint64(x[len(key):], counter)
	endian.PutUint16(x[len(key)+8:], uint16(len(key)))

	return x
}

type keyData struct {
	// key is the key unmodified.
	key []byte
	// prefix is the key prefix.
	prefix []byte
	// suffix is the suffix, which is a uint64.
	suffix []byte
}

// getPrefixSuffix assumes that you are pulling a key that was encoded with a suffix, aka
// a key in an allow duplicates index. If not, the key data is going to error or just be bad.
func getPrefixSuffix(key []byte) (keyData, error) {
	if len(key) < 3 { // 1 byte + uint16
		return keyData{}, fmt.Errorf("getPrefixSuffix encountered key of size: %d", len(key))
	}
	if len(key) >= math.MaxUint16 {
		return keyData{}, fmt.Errorf("cannot have a key > %d bytes", math.MaxUint16)
	}

	prefixSize := endian.Uint16(key[len(key)-2:])
	if int(prefixSize) > len(key)-2 {
		return keyData{}, fmt.Errorf("getPrefixSuffix encountered key with prefix size greater than the key length")
	}

	prefix := key[:prefixSize]
	var suffix []byte
	if len(key)-2 > int(prefixSize) {
		suffix = key[prefixSize : len(key)-2]
	}

	if suffix == nil {
		if len(key) != 2+int(prefixSize) {
			return keyData{}, fmt.Errorf("key is corrupt, we have a uint16 footer that lists the prefix of size %d and a suffix of uint64, but have %d bytes", prefixSize, len(key))
		}
	} else {
		if len(key) != 2+int(prefixSize)+8 {
			return keyData{}, fmt.Errorf("key is corrupt, we have a uint16 footer that lists the prefix of size %d and a suffix of uint64, but have %d bytes", prefixSize, len(key))
		}
	}

	return keyData{key: key, prefix: prefix, suffix: suffix}, nil
}

// nullLogger is used to log all the badger output to null.
type nullLogger struct{}

func (n nullLogger) Errorf(string, ...interface{})   {}
func (n nullLogger) Warningf(string, ...interface{}) {}
func (n nullLogger) Infof(string, ...interface{})    {}
func (n nullLogger) Debugf(string, ...interface{})   {}