//go:build kvdb_etcd
// +build kvdb_etcd

package etcd

import (
	"context"
	"fmt"
	"math"
	"strings"

	"github.com/google/btree"
	pb "go.etcd.io/etcd/api/v3/etcdserverpb"
	v3 "go.etcd.io/etcd/client/v3"
)

type CommitStats struct {
	Rset    int
	Wset    int
	Retries int
}

// KV stores a key/value pair.
type KV struct {
	key string
	val string
}

// STM is an interface for software transactional memory.
// All calls that return error will do so only if STM is manually handled and
// abort the apply closure otherwise. In both case the returned error is a
// DatabaseError.
type STM interface {
	// Get returns the value for a key and inserts the key in the txn's read
	// set. Returns nil if there's no matching key, or the key is empty.
	Get(key string) ([]byte, error)

	// Put adds a value for a key to the txn's write set.
	Put(key, val string)

	// Del adds a delete operation for the key to the txn's write set.
	Del(key string)

	// First returns the first k/v that begins with prefix or nil if there's
	// no such k/v pair. If the key is found it is inserted to the txn's
	// read set. Returns nil if there's no match.
	First(prefix string) (*KV, error)

	// Last returns the last k/v that begins with prefix or nil if there's
	// no such k/v pair. If the key is found it is inserted to the txn's
	// read set. Returns nil if there's no match.
	Last(prefix string) (*KV, error)

	// Prev returns the previous k/v before key that begins with prefix or
	// nil if there's no such k/v. If the key is found it is inserted to the
	// read set. Returns nil if there's no match.
	Prev(prefix, key string) (*KV, error)

	// Next returns the next k/v after key that begins with prefix or nil
	// if there's no such k/v. If the key is found it is inserted to the
	// txn's read set. Returns nil if there's no match.
	Next(prefix, key string) (*KV, error)

	// Seek will return k/v at key beginning with prefix. If the key doesn't
	// exists Seek will return the next k/v after key beginning with prefix.
	// If a matching k/v is found it is inserted to the txn's read set. Returns
	// nil if there's no match.
	Seek(prefix, key string) (*KV, error)

	// OnCommit calls the passed callback func upon commit.
	OnCommit(func())

	// Commit attempts to apply the txn's changes to the server.
	// Commit may return CommitError if transaction is outdated and needs retry.
	Commit() error

	// Rollback entries the read and write sets such that a subsequent commit
	// won't alter the database.
	Rollback()

	// Prefetch prefetches the passed keys and prefixes. For prefixes it'll
	// fetch the whole range.
	Prefetch(keys []string, prefix []string)

	// FetchRangePaginatedRaw will fetch the range with the passed prefix up
	// to the passed limit per page.
	FetchRangePaginatedRaw(prefix string, limit int64,
		cb func(kv KV) error) error
}

// CommitError is used to check if there was an error
// due to stale data in the transaction.
type CommitError struct{}

// Error returns a static string for CommitError for
// debugging/logging purposes.
func (e CommitError) Error() string {
	return "commit failed"
}

// DatabaseError is used to wrap errors that are not
// related to stale data in the transaction.
type DatabaseError struct {
	msg string
	err error
}

// Unwrap returns the wrapped error in a DatabaseError.
func (e *DatabaseError) Unwrap() error {
	return e.err
}

// Error simply converts DatabaseError to a string that
// includes both the message and the wrapped error.
func (e DatabaseError) Error() string {
	return fmt.Sprintf("etcd error: %v - %v", e.msg, e.err)
}

// stmGet is the result of a read operation, a value and the mod revision of the
// key/value.
type stmGet struct {
	KV
	rev int64
}

// Less implements less operator for btree.BTree.
func (c *stmGet) Less(than btree.Item) bool {
	return c.key < than.(*stmGet).key
}

// readSet stores all reads done in an STM.
type readSet struct {
	// tree stores the items in the read set.
	tree *btree.BTree

	// fullRanges stores full range prefixes.
	fullRanges map[string]struct{}
}

// stmPut stores a value and an operation (put/delete).
type stmPut struct {
	val string
	op  v3.Op
}

// writeSet stroes all writes done in an STM.
type writeSet map[string]stmPut

// stm implements repeatable-read software transactional memory
// over etcd.
type stm struct {
	// client is an etcd client handling all RPC communications
	// to the etcd instance/cluster.
	client *v3.Client

	// manual is set to true for manual transactions which don't
	// execute in the STM run loop.
	manual bool

	// txQueue is lightweight contention manager, which is used to detect
	// transaction conflicts and reduce retries.
	txQueue *commitQueue

	// options stores optional settings passed by the user.
	options *STMOptions

	// rset holds read key values and revisions.
	rset *readSet

	// wset holds overwritten keys and their values.
	wset writeSet

	// getOpts are the opts used for gets.
	getOpts []v3.OpOption

	// revision stores the snapshot revision after first read.
	revision int64

	// onCommit gets called upon commit.
	onCommit func()

	// callCount tracks the number of times we called into etcd.
	callCount int
}

// STMOptions can be used to pass optional settings
// when an STM is created.
type STMOptions struct {
	// ctx holds an externally provided abort context.
	ctx                 context.Context
	commitStatsCallback func(bool, CommitStats)
}

// STMOptionFunc is a function that updates the passed STMOptions.
type STMOptionFunc func(*STMOptions)

// WithAbortContext specifies the context for permanently
// aborting the transaction.
func WithAbortContext(ctx context.Context) STMOptionFunc {
	return func(so *STMOptions) {
		so.ctx = ctx
	}
}

func WithCommitStatsCallback(cb func(bool, CommitStats)) STMOptionFunc {
	return func(so *STMOptions) {
		so.commitStatsCallback = cb
	}
}

// RunSTM runs the apply function by creating an STM using serializable snapshot
// isolation, passing it to the apply and handling commit errors and retries.
func RunSTM(cli *v3.Client, apply func(STM) error, txQueue *commitQueue,
	so ...STMOptionFunc) (int, error) {

	stm := makeSTM(cli, false, txQueue, so...)
	err := runSTM(stm, apply)

	return stm.callCount, err
}

// NewSTM creates a new STM instance, using serializable snapshot isolation.
func NewSTM(cli *v3.Client, txQueue *commitQueue, so ...STMOptionFunc) STM {
	return makeSTM(cli, true, txQueue, so...)
}

// makeSTM is the actual constructor of the stm. It first apply all passed
// options then creates the stm object and resets it before returning.
func makeSTM(cli *v3.Client, manual bool, txQueue *commitQueue,
	so ...STMOptionFunc) *stm {

	opts := &STMOptions{
		ctx: cli.Ctx(),
	}

	// Apply all functional options.
	for _, fo := range so {
		fo(opts)
	}

	s := &stm{
		client:  cli,
		manual:  manual,
		txQueue: txQueue,
		options: opts,
		rset:    newReadSet(),
	}

	// Reset read and write set.
	s.rollback(true)

	return s
}

// runSTM implements the run loop of the STM, running the apply func, catching
// errors and handling commit. The loop will quit on every error except
// CommitError which is used to indicate a necessary retry.
func runSTM(s *stm, apply func(STM) error) error {
	var (
		retries    int
		stats      CommitStats
		executeErr error
	)

	done := make(chan struct{})

	execute := func() {
		defer close(done)

		for {
			select {
			// Check if the STM is aborted and break the retry loop
			// if it is.
			case <-s.options.ctx.Done():
				executeErr = fmt.Errorf("aborted")
				return

			default:
			}

			stats, executeErr = s.commit()

			// Re-apply only upon commit error (meaning the
			// keys were changed).
			if _, ok := executeErr.(CommitError); !ok {
				// Anything that's not a CommitError
				// aborts the transaction.
				return
			}

			// Rollback the write set before trying to re-apply.
			// Upon commit we retrieved the latest version of all
			// previously fetched keys and ranges so we don't need
			// to rollback the read set.
			s.rollback(false)
			retries++

			// Re-apply the transaction closure.
			if executeErr = apply(s); executeErr != nil {
				return
			}
		}
	}

	// Run the tx closure to construct the read and write sets.
	// Also we expect that if there are no conflicting transactions
	// in the queue, then we only run apply once.
	if preApplyErr := apply(s); preApplyErr != nil {
		return preApplyErr
	}

	// Make a copy of the read/write set keys here. The reason why we need
	// to do this is because subsequent applies may change (shrink) these
	// sets and so when we decrease reference counts in the commit queue in
	// done(...) we'd potentially miss removing references which would
	// result in queueing up transactions and contending DB access.
	// Copying these strings is cheap due to Go's immutable string which is
	// always a reference.
	rkeys := make([]string, s.rset.tree.Len())
	wkeys := make([]string, len(s.wset))

	i := 0
	s.rset.tree.Ascend(func(item btree.Item) bool {
		rkeys[i] = item.(*stmGet).key
		i++

		return true
	})

	i = 0
	for key := range s.wset {
		wkeys[i] = key
		i++
	}

	// Queue up the transaction for execution.
	s.txQueue.Add(execute, rkeys, wkeys)

	// Wait for the transaction to execute, or break if aborted.
	select {
	case <-done:
	case <-s.options.ctx.Done():
		return context.Canceled
	}

	if s.options.commitStatsCallback != nil {
		stats.Retries = retries
		s.options.commitStatsCallback(executeErr == nil, stats)
	}

	return executeErr
}

func newReadSet() *readSet {
	return &readSet{
		tree:       btree.New(5),
		fullRanges: make(map[string]struct{}),
	}
}

// add inserts key/values to to read set.
func (rs *readSet) add(responses []*pb.ResponseOp) {
	for _, resp := range responses {
		getResp := resp.GetResponseRange()
		for _, kv := range getResp.Kvs {
			rs.addItem(
				string(kv.Key), string(kv.Value), kv.ModRevision,
			)
		}
	}
}

// addFullRange adds all full ranges to the read set.
func (rs *readSet) addFullRange(prefixes []string, responses []*pb.ResponseOp) {
	for i, resp := range responses {
		getResp := resp.GetResponseRange()
		for _, kv := range getResp.Kvs {
			rs.addItem(
				string(kv.Key), string(kv.Value), kv.ModRevision,
			)
		}

		rs.fullRanges[prefixes[i]] = struct{}{}
	}
}

// presetItem presets a key to zero revision if not already present in the read
// set.
func (rs *readSet) presetItem(key string) {
	item := &stmGet{
		KV: KV{
			key: key,
		},
		rev: 0,
	}

	if !rs.tree.Has(item) {
		rs.tree.ReplaceOrInsert(item)
	}
}

// addItem adds a single new key/value to the read set (if not already present).
func (rs *readSet) addItem(key, val string, modRevision int64) {
	item := &stmGet{
		KV: KV{
			key: key,
			val: val,
		},
		rev: modRevision,
	}

	rs.tree.ReplaceOrInsert(item)
}

// hasFullRange checks if the read set has a full range prefetched.
func (rs *readSet) hasFullRange(prefix string) bool {
	_, ok := rs.fullRanges[prefix]
	return ok
}

// next returns the pre-fetched next value of the prefix. If matchKey is true,
// it'll simply return the key/value that matches the passed key.
func (rs *readSet) next(prefix, key string, matchKey bool) (*stmGet, bool) {
	pivot := &stmGet{
		KV: KV{
			key: key,
		},
	}

	var result *stmGet
	rs.tree.AscendGreaterOrEqual(
		pivot,
		func(item btree.Item) bool {
			next := item.(*stmGet)
			if (!matchKey && next.key == key) || next.rev == 0 {
				return true
			}

			if strings.HasPrefix(next.key, prefix) {
				result = next
			}

			return false
		},
	)

	return result, result != nil
}

// prev returns the pre-fetched prev key/value of the prefix from key.
func (rs *readSet) prev(prefix, key string) (*stmGet, bool) {
	pivot := &stmGet{
		KV: KV{
			key: key,
		},
	}

	var result *stmGet

	rs.tree.DescendLessOrEqual(
		pivot, func(item btree.Item) bool {
			prev := item.(*stmGet)
			if prev.key == key || prev.rev == 0 {
				return true
			}

			if strings.HasPrefix(prev.key, prefix) {
				result = prev
			}

			return false
		},
	)

	return result, result != nil
}

// last returns the last key/value of the passed range (if prefetched).
func (rs *readSet) last(prefix string) (*stmGet, bool) {
	// We create an artificial key here that is just one step away from the
	// prefix. This way when we try to get the first item with our prefix
	// before this newly crafted key we'll make sure it's the last element
	// of our range.
	key := []byte(prefix)
	key[len(key)-1] += 1

	return rs.prev(prefix, string(key))
}

// clear completely clears the readset.
func (rs *readSet) clear() {
	rs.tree.Clear(false)
	rs.fullRanges = make(map[string]struct{})
}

// getItem returns the matching key/value from the readset.
func (rs *readSet) getItem(key string) (*stmGet, bool) {
	pivot := &stmGet{
		KV: KV{
			key: key,
		},
		rev: 0,
	}
	item := rs.tree.Get(pivot)
	if item != nil {
		return item.(*stmGet), true
	}

	// It's possible that although this key isn't in the read set, we
	// fetched a full range the key is prefixed with. In this case we'll
	// insert the key with zero revision.
	for prefix := range rs.fullRanges {
		if strings.HasPrefix(key, prefix) {
			rs.tree.ReplaceOrInsert(pivot)
			return pivot, true
		}
	}

	return nil, false
}

// prefetchSet is a helper to create an op slice of all OpGet's that represent
// fetched keys appended with a slice of all OpGet's representing all prefetched
// full ranges.
func (rs *readSet) prefetchSet() []v3.Op {
	ops := make([]v3.Op, 0, rs.tree.Len())

	rs.tree.Ascend(func(item btree.Item) bool {
		key := item.(*stmGet).key
		for prefix := range rs.fullRanges {
			// Do not add the key if it has been prefetched in a
			// full range.
			if strings.HasPrefix(key, prefix) {
				return true
			}
		}

		ops = append(ops, v3.OpGet(key))
		return true
	})

	for prefix := range rs.fullRanges {
		ops = append(ops, v3.OpGet(prefix, v3.WithPrefix()))
	}

	return ops
}

// getFullRanges returns all prefixes that we prefetched.
func (rs *readSet) getFullRanges() []string {
	prefixes := make([]string, 0, len(rs.fullRanges))

	for prefix := range rs.fullRanges {
		prefixes = append(prefixes, prefix)
	}

	return prefixes
}

// cmps returns a compare list which will serve as a precondition testing that
// the values in the read set didn't change.
func (rs *readSet) cmps() []v3.Cmp {
	cmps := make([]v3.Cmp, 0, rs.tree.Len())

	rs.tree.Ascend(func(item btree.Item) bool {
		get := item.(*stmGet)
		cmps = append(
			cmps, v3.Compare(v3.ModRevision(get.key), "=", get.rev),
		)

		return true
	})

	return cmps
}

// cmps returns a cmp list testing no writes have happened past rev.
func (ws writeSet) cmps(rev int64) []v3.Cmp {
	cmps := make([]v3.Cmp, 0, len(ws))
	for key := range ws {
		cmps = append(cmps, v3.Compare(v3.ModRevision(key), "<", rev))
	}

	return cmps
}

// puts is the list of ops for all pending writes.
func (ws writeSet) puts() []v3.Op {
	puts := make([]v3.Op, 0, len(ws))
	for _, v := range ws {
		puts = append(puts, v.op)
	}

	return puts
}

// FetchRangePaginatedRaw will fetch the range with the passed prefix up to the
// passed limit per page.
func (s *stm) FetchRangePaginatedRaw(prefix string, limit int64,
	cb func(kv KV) error) error {

	s.callCount++

	opts := []v3.OpOption{
		v3.WithSort(v3.SortByKey, v3.SortAscend),
		v3.WithRange(v3.GetPrefixRangeEnd(prefix)),
		v3.WithLimit(limit),
	}

	key := prefix
	for {
		resp, err := s.client.Get(
			s.options.ctx, key, append(opts, s.getOpts...)...,
		)
		if err != nil {
			return DatabaseError{
				msg: "stm.fetch() failed",
				err: err,
			}
		}

		// Fill the read set with key/values returned.
		for _, kv := range resp.Kvs {
			err := cb(KV{string(kv.Key), string(kv.Value)})

			if err != nil {
				return err
			}
		}

		// We've reached the range end.
		if !resp.More {
			break
		}

		// Continue from the page end + "\x00".
		key = string(resp.Kvs[len(resp.Kvs)-1].Key) + "\x00"
	}

	return nil
}

// fetch is a helper to fetch key/value given options. If a value is returned
// then fetch will try to fix the STM's snapshot revision (if not already set).
// We'll also cache the returned key/value in the read set.
func (s *stm) fetch(key string, opts ...v3.OpOption) ([]KV, error) {
	s.callCount++
	resp, err := s.client.Get(
		s.options.ctx, key, append(opts, s.getOpts...)...,
	)
	if err != nil {
		return nil, DatabaseError{
			msg: "stm.fetch() failed",
			err: err,
		}
	}

	// Set revision and serializable options upon first fetch
	// for any subsequent fetches.
	if s.getOpts == nil {
		s.revision = resp.Header.Revision
		s.getOpts = []v3.OpOption{
			v3.WithRev(s.revision),
			v3.WithSerializable(),
		}
	}

	if len(resp.Kvs) == 0 {
		// Add assertion to the read set which will extend our commit
		// constraint such that the commit will fail if the key is
		// present in the database.
		s.rset.addItem(key, "", 0)
	}

	var result []KV

	// Fill the read set with key/values returned.
	for _, kv := range resp.Kvs {
		key := string(kv.Key)
		val := string(kv.Value)

		// Add to read set.
		s.rset.addItem(key, val, kv.ModRevision)

		result = append(result, KV{key, val})
	}

	return result, nil
}

// Get returns the value for key. If there's no such
// key/value in the database or the passed key is empty
// Get will return nil.
func (s *stm) Get(key string) ([]byte, error) {
	if key == "" {
		return nil, nil
	}

	// Return freshly written value if present.
	if put, ok := s.wset[key]; ok {
		if put.op.IsDelete() {
			return nil, nil
		}

		return []byte(put.val), nil
	}

	// Return value if alread in read set.
	if getValue, ok := s.rset.getItem(key); ok {
		// Return the value if the rset contains an existing key.
		if getValue.rev != 0 {
			return []byte(getValue.val), nil
		} else {
			return nil, nil
		}
	}

	// Fetch and return value.
	kvs, err := s.fetch(key)
	if err != nil {
		return nil, err
	}

	if len(kvs) > 0 {
		return []byte(kvs[0].val), nil
	}

	// Return empty result if key not in DB.
	return nil, nil
}

// First returns the first key/value matching prefix. If there's no key starting
// with prefix, Last will return nil.
func (s *stm) First(prefix string) (*KV, error) {
	return s.next(prefix, prefix, true)
}

// Last returns the last key/value with prefix. If there's no key starting with
// prefix, Last will return nil.
func (s *stm) Last(prefix string) (*KV, error) {
	var (
		kv    KV
		found bool
	)

	if s.rset.hasFullRange(prefix) {
		if item, ok := s.rset.last(prefix); ok {
			kv = item.KV
			found = true
		}
	} else {
		// As we don't know the full range, fetch the last
		// key/value with this prefix first.
		resp, err := s.fetch(prefix, v3.WithLastKey()...)
		if err != nil {
			return nil, err
		}

		if len(resp) > 0 {
			kv = resp[0]
			found = true
		}
	}

	// Now make sure there's nothing in the write set
	// that is a better match, meaning it has the same
	// prefix but is greater or equal than the current
	// best candidate. Note that this is not efficient
	// when the write set is large!
	for k, put := range s.wset {
		if put.op.IsDelete() {
			continue
		}

		if strings.HasPrefix(k, prefix) && k >= kv.key {
			kv.key = k
			kv.val = put.val
			found = true
		}
	}

	if found {
		return &kv, nil
	}

	return nil, nil
}

// Prev returns the prior key/value before key (with prefix). If there's no such
// key Prev will return nil.
func (s *stm) Prev(prefix, startKey string) (*KV, error) {
	var kv, result KV

	fetchKey := startKey
	matchFound := false

	for {
		if s.rset.hasFullRange(prefix) {
			if item, ok := s.rset.prev(prefix, fetchKey); ok {
				kv = item.KV
			} else {
				break
			}
		} else {

			// Ask etcd to retrieve one key that is a
			// match in descending order from the passed key.
			opts := []v3.OpOption{
				v3.WithRange(fetchKey),
				v3.WithSort(v3.SortByKey, v3.SortDescend),
				v3.WithLimit(1),
			}

			kvs, err := s.fetch(prefix, opts...)
			if err != nil {
				return nil, err
			}

			if len(kvs) == 0 {
				break
			}

			kv = kvs[0]
		}

		// WithRange and WithPrefix can't be used
		// together, so check prefix here. If the
		// returned key no longer has the prefix,
		// then break out.
		if !strings.HasPrefix(kv.key, prefix) {
			break
		}

		// Fetch the prior key if this is deleted.
		if put, ok := s.wset[kv.key]; ok && put.op.IsDelete() {
			fetchKey = kv.key
			continue
		}

		result = kv
		matchFound = true

		break
	}

	// Closure holding all checks to find a possibly
	// better match.
	matches := func(key string) bool {
		if !strings.HasPrefix(key, prefix) {
			return false
		}

		if !matchFound {
			return key < startKey
		}

		// matchFound == true
		return result.key <= key && key < startKey
	}

	// Now go trough the write set and check
	// if there's an even better match.
	for k, put := range s.wset {
		if !put.op.IsDelete() && matches(k) {
			result.key = k
			result.val = put.val
			matchFound = true
		}
	}

	if !matchFound {
		return nil, nil
	}

	return &result, nil
}

// Next returns the next key/value after key (with prefix). If there's no such
// key Next will return nil.
func (s *stm) Next(prefix string, key string) (*KV, error) {
	return s.next(prefix, key, false)
}

// Seek "seeks" to the key (with prefix). If the key doesn't exists it'll get
// the next key with the same prefix. If no key fills this criteria, Seek will
// return nil.
func (s *stm) Seek(prefix, key string) (*KV, error) {
	return s.next(prefix, key, true)
}

// next will try to retrieve the next match that has prefix and starts with the
// passed startKey. If includeStartKey is set to true, it'll return the value
// of startKey (essentially implementing seek).
func (s *stm) next(prefix, startKey string, includeStartKey bool) (*KV, error) {
	var kv, result KV

	fetchKey := startKey
	firstFetch := true
	matchFound := false

	for {
		if s.rset.hasFullRange(prefix) {
			matchKey := includeStartKey && firstFetch
			firstFetch = false
			if item, ok := s.rset.next(
				prefix, fetchKey, matchKey,
			); ok {
				kv = item.KV
			} else {
				break
			}
		} else {
			// Ask etcd to retrieve one key that is a
			// match in ascending order from the passed key.
			opts := []v3.OpOption{
				v3.WithFromKey(),
				v3.WithSort(v3.SortByKey, v3.SortAscend),
				v3.WithLimit(1),
			}

			// By default we include the start key too
			// if it is a full match.
			if includeStartKey && firstFetch {
				firstFetch = false
			} else {
				// If we'd like to retrieve the first key
				// after the start key.
				fetchKey += "\x00"
			}

			kvs, err := s.fetch(fetchKey, opts...)
			if err != nil {
				return nil, err
			}

			if len(kvs) == 0 {
				break
			}

			kv = kvs[0]

			// WithRange and WithPrefix can't be used
			// together, so check prefix here. If the
			// returned key no longer has the prefix,
			// then break the fetch loop.
			if !strings.HasPrefix(kv.key, prefix) {
				break
			}
		}

		// Move on to fetch starting with the next
		// key if this one is marked deleted.
		if put, ok := s.wset[kv.key]; ok && put.op.IsDelete() {
			fetchKey = kv.key
			continue
		}

		result = kv
		matchFound = true

		break
	}

	// Closure holding all checks to find a possibly
	// better match.
	matches := func(k string) bool {
		if !strings.HasPrefix(k, prefix) {
			return false
		}

		if includeStartKey && !matchFound {
			return startKey <= k
		}

		if !includeStartKey && !matchFound {
			return startKey < k
		}

		if includeStartKey && matchFound {
			return startKey <= k && k <= result.key
		}

		// !includeStartKey && matchFound.
		return startKey < k && k <= result.key
	}

	// Now go trough the write set and check
	// if there's an even better match.
	for k, put := range s.wset {
		if !put.op.IsDelete() && matches(k) {
			result.key = k
			result.val = put.val
			matchFound = true
		}
	}

	if !matchFound {
		return nil, nil
	}

	return &result, nil
}

// Put sets the value of the passed key. The actual put will happen upon commit.
func (s *stm) Put(key, val string) {
	s.wset[key] = stmPut{
		val: val,
		op:  v3.OpPut(key, val),
	}
}

// Del marks a key as deleted. The actual delete will happen upon commit.
func (s *stm) Del(key string) {
	s.wset[key] = stmPut{
		val: "",
		op:  v3.OpDelete(key),
	}
}

// OnCommit sets the callback that is called upon committing the STM
// transaction.
func (s *stm) OnCommit(cb func()) {
	s.onCommit = cb
}

// Prefetch will prefetch the passed keys and prefixes in one transaction.
// Keys and prefixes that we already have will be skipped.
func (s *stm) Prefetch(keys []string, prefixes []string) {
	fetchKeys := make([]string, 0, len(keys))
	for _, key := range keys {
		if _, ok := s.rset.getItem(key); !ok {
			fetchKeys = append(fetchKeys, key)
		}
	}

	fetchPrefixes := make([]string, 0, len(prefixes))
	for _, prefix := range prefixes {
		if s.rset.hasFullRange(prefix) {
			continue
		}
		fetchPrefixes = append(fetchPrefixes, prefix)
	}

	if len(fetchKeys) == 0 && len(fetchPrefixes) == 0 {
		return
	}

	prefixOpts := append(
		[]v3.OpOption{v3.WithPrefix()}, s.getOpts...,
	)

	txn := s.client.Txn(s.options.ctx)
	ops := make([]v3.Op, 0, len(fetchKeys)+len(fetchPrefixes))

	for _, key := range fetchKeys {
		ops = append(ops, v3.OpGet(key, s.getOpts...))
	}
	for _, key := range fetchPrefixes {
		ops = append(ops, v3.OpGet(key, prefixOpts...))
	}

	txn.Then(ops...)
	txnresp, err := txn.Commit()
	s.callCount++

	if err != nil {
		return
	}

	// Set revision and serializable options upon first fetch for any
	// subsequent fetches.
	if s.getOpts == nil {
		s.revision = txnresp.Header.Revision
		s.getOpts = []v3.OpOption{
			v3.WithRev(s.revision),
			v3.WithSerializable(),
		}
	}

	// Preset keys to "not-present" (revision set to zero).
	for _, key := range fetchKeys {
		s.rset.presetItem(key)
	}

	// Set prefetched keys.
	s.rset.add(txnresp.Responses[:len(fetchKeys)])

	// Set prefetched ranges.
	s.rset.addFullRange(fetchPrefixes, txnresp.Responses[len(fetchKeys):])
}

// commit builds the final transaction and tries to execute it. If commit fails
// because the keys have changed return a CommitError, otherwise return a
// DatabaseError.
func (s *stm) commit() (CommitStats, error) {
	rset := s.rset.cmps()
	wset := s.wset.cmps(s.revision + 1)

	stats := CommitStats{
		Rset: len(rset),
		Wset: len(wset),
	}

	// Create the compare set.
	cmps := append(rset, wset...)
	// Create a transaction with the optional abort context.
	txn := s.client.Txn(s.options.ctx)

	// If the compare set holds, try executing the puts.
	txn = txn.If(cmps...)
	txn = txn.Then(s.wset.puts()...)

	// Prefetch keys and ranges in case of conflict to save as many
	// round-trips as possible.
	txn = txn.Else(s.rset.prefetchSet()...)

	s.callCount++
	txnresp, err := txn.Commit()
	if err != nil {
		return stats, DatabaseError{
			msg: "stm.Commit() failed",
			err: err,
		}
	}

	// Call the commit callback if the transaction was successful.
	if txnresp.Succeeded {
		if s.onCommit != nil {
			s.onCommit()
		}

		return stats, nil
	}

	// Determine where our fetched full ranges begin in the response.
	prefixes := s.rset.getFullRanges()
	firstPrefixResp := len(txnresp.Responses) - len(prefixes)

	// Clear reload and preload it with the prefetched keys and ranges.
	s.rset.clear()
	s.rset.add(txnresp.Responses[:firstPrefixResp])
	s.rset.addFullRange(prefixes, txnresp.Responses[firstPrefixResp:])

	// Set our revision boundary.
	s.revision = txnresp.Header.Revision
	s.getOpts = []v3.OpOption{
		v3.WithRev(s.revision),
		v3.WithSerializable(),
	}

	// Return CommitError indicating that the transaction can be retried.
	return stats, CommitError{}
}

// Commit simply calls commit and the commit stats callback if set.
func (s *stm) Commit() error {
	stats, err := s.commit()

	if s.options.commitStatsCallback != nil {
		s.options.commitStatsCallback(err == nil, stats)
	}

	return err
}

// Rollback resets the STM. This is useful for uncommitted transaction rollback
// and also used in the STM main loop to reset state if commit fails.
func (s *stm) Rollback() {
	s.rollback(true)
}

// rollback will reset the read and write sets. If clearReadSet is false we'll
// only reset the the write set.
func (s *stm) rollback(clearReadSet bool) {
	if clearReadSet {
		s.rset.clear()
		s.revision = math.MaxInt64 - 1
		s.getOpts = nil
	}

	s.wset = make(map[string]stmPut)
}