feat!: Refactor PrepareProposal to produce blocks using the non-int…

…eractive defaults (#692)

Co-authored-by: Rootul Patel <[email protected]>

app/estimate_square_size.go

@@ -0,0 +1,244 @@
+package app
+
+import (
+	"bytes"
+	"math"
+	"sort"
+
+	"github.com/celestiaorg/celestia-app/pkg/appconsts"
+	"github.com/celestiaorg/celestia-app/pkg/shares"
+	"github.com/cosmos/cosmos-sdk/client"
+	core "github.com/tendermint/tendermint/proto/tendermint/types"
+	coretypes "github.com/tendermint/tendermint/types"
+)
+
+// prune removes txs until the set of txs will fit in the square of size
+// squareSize. It assumes that the currentShareCount is accurate. This function
+// is far from optimal because accurately knowing how many shares any given
+// set of transactions and its message takes up in a data square that is following the
+// non-interactive default rules requires recalculating the entire square.
+// TODO: include the padding used by each msg when counting removed shares
+func prune(txConf client.TxConfig, txs []*parsedTx, currentShareCount, squareSize int) parsedTxs {
+	maxShares := squareSize * squareSize
+	if maxShares >= currentShareCount {
+		return txs
+	}
+	goal := currentShareCount - maxShares
+
+	removedContiguousShares := 0
+	contigBytesCursor := 0
+	removedMessageShares := 0
+	removedTxs := 0
+
+	// adjustContigCursor checks if enough contiguous bytes have been removed
+	// inorder to tally total contiguous shares removed
+	adjustContigCursor := func(l int) {
+		contigBytesCursor += l + shares.DelimLen(uint64(l))
+		if contigBytesCursor >= appconsts.CompactShareContentSize {
+			removedContiguousShares += (contigBytesCursor / appconsts.CompactShareContentSize)
+			contigBytesCursor = contigBytesCursor % appconsts.CompactShareContentSize
+		}
+	}
+
+	for i := len(txs) - 1; (removedContiguousShares + removedMessageShares) < goal; i-- {
+		// this normally doesn't happen, but since we don't calculate the number
+		// of padded shares also being removed, its possible to reach this value
+		// should there be many small messages, and we don't want to panic.
+		if i < 0 {
+			break
+		}
+		removedTxs++
+		if txs[i].msg == nil {
+			adjustContigCursor(len(txs[i].rawTx))
+			continue
+		}
+
+		removedMessageShares += shares.MsgSharesUsed(len(txs[i].msg.GetMessage()))
+		// we ignore the error here, as if there is an error malleating the tx,
+		// then we need to remove it anyway and it will not end up contributing
+		// bytes to the square anyway.
+		_ = txs[i].malleate(txConf, uint64(squareSize))
+		adjustContigCursor(len(txs[i].malleatedTx) + appconsts.MalleatedTxBytes)
+	}
+
+	return txs[:len(txs)-(removedTxs)]
+}
+
+// calculateCompactShareCount calculates the exact number of compact shares used.
+func calculateCompactShareCount(txs []*parsedTx, evd core.EvidenceList, squareSize int) int {
+	txSplitter := shares.NewCompactShareSplitter(appconsts.TxNamespaceID, appconsts.ShareVersion)
+	evdSplitter := shares.NewCompactShareSplitter(appconsts.EvidenceNamespaceID, appconsts.ShareVersion)
+	var err error
+	msgSharesCursor := len(txs)
+	for _, tx := range txs {
+		rawTx := tx.rawTx
+		if tx.malleatedTx != nil {
+			rawTx, err = coretypes.WrapMalleatedTx(tx.originalHash(), uint32(msgSharesCursor), tx.malleatedTx)
+			if err != nil {
+				panic(err)
+			}
+			used, _ := shares.MsgSharesUsedNonInteractiveDefaults(msgSharesCursor, squareSize, tx.msg.Size())
+			msgSharesCursor += used
+		}
+		txSplitter.WriteTx(rawTx)
+	}
+	for _, e := range evd.Evidence {
+		evidence, err := coretypes.EvidenceFromProto(&e)
+		if err != nil {
+			panic(err)
+		}
+		err = evdSplitter.WriteEvidence(evidence)
+		if err != nil {
+			panic(err)
+		}
+	}
+	txCount, available := txSplitter.Count()
+	if appconsts.CompactShareContentSize-available > 0 {
+		txCount++
+	}
+	evdCount, available := evdSplitter.Count()
+	if appconsts.CompactShareContentSize-available > 0 {
+		evdCount++
+	}
+	return txCount + evdCount
+}
+
+// estimateSquareSize uses the provided block data to estimate the square size
+// assuming that all malleated txs follow the non interactive default rules.
+// Returns the estimated square size and the number of shares used.
+func estimateSquareSize(txs []*parsedTx, evd core.EvidenceList) (uint64, int) {
+	// get the raw count of shares taken by each type of block data
+	txShares, evdShares, msgLens := rawShareCount(txs, evd)
+	msgShares := 0
+	for _, msgLen := range msgLens {
+		msgShares += msgLen
+	}
+
+	// calculate the smallest possible square size that could contain all the
+	// messages
+	squareSize := nextPowerOfTwo(int(math.Ceil(math.Sqrt(float64(txShares + evdShares + msgShares)))))
+
+	// the starting square size should at least be the minimum
+	if squareSize < appconsts.MinSquareSize {
+		squareSize = appconsts.MinSquareSize
+	}
+
+	var fits bool
+	for {
+		// assume that all the msgs in the square use the non-interactive
+		// default rules and see if we can fit them in the smallest starting
+		// square size. We start the cursor (share index) at the beginning of
+		// the message shares (txShares+evdShares), because shares that do not
+		// follow the non-interactive defaults are simple to estimate.
+		fits, msgShares = shares.FitsInSquare(txShares+evdShares, squareSize, msgLens...)
+		switch {
+		// stop estimating if we know we can reach the max square size
+		case squareSize >= appconsts.MaxSquareSize:
+			return appconsts.MaxSquareSize, txShares + evdShares + msgShares
+		// return if we've found a square size that fits all of the txs
+		case fits:
+			return uint64(squareSize), txShares + evdShares + msgShares
+		// try the next largest square size if we can't fit all the txs
+		case !fits:
+			// double the square size
+			squareSize = nextPowerOfTwo(squareSize + 1)
+		}
+	}
+}
+
+// rawShareCount calculates the number of shares taken by all of the included
+// txs, evidence, and each msg. msgLens is a slice of the number of shares used
+// by each message without accounting for the non-interactive message layout
+// rules.
+func rawShareCount(txs []*parsedTx, evd core.EvidenceList) (txShares, evdShares int, msgLens []int) {
+	// msgSummary is used to keep track of the size and the namespace so that we
+	// can sort the messages by namespace before returning.
+	type msgSummary struct {
+		// size is the number of shares used by this message
+		size      int
+		namespace []byte
+	}
+
+	var msgSummaries []msgSummary
+
+	// we use bytes instead of shares for tx and evd as they are encoded
+	// contiguously in the square, unlike msgs where each of which is assigned their
+	// own set of shares
+	txBytes, evdBytes := 0, 0
+	for _, pTx := range txs {
+		// if there is no wire message in this tx, then we can simply add the
+		// bytes and move on.
+		if pTx.msg == nil {
+			txBytes += len(pTx.rawTx)
+			continue
+		}
+
+		// if there is a malleated tx, then we want to also account for the
+		// txs that get included on-chain. The formula used here over
+		// compensates for the actual size of the message, and in some cases can
+		// result in some wasted square space or picking a square size that is
+		// too large. TODO: improve by making a more accurate estimation formula
+		txBytes += overEstimateMalleatedTxSize(len(pTx.rawTx), len(pTx.msg.Message), len(pTx.msg.MessageShareCommitment))
+
+		msgSummaries = append(msgSummaries, msgSummary{shares.MsgSharesUsed(int(pTx.msg.MessageSize)), pTx.msg.MessageNameSpaceId})
+	}
+
+	txShares = txBytes / appconsts.CompactShareContentSize
+	if txBytes > 0 {
+		txShares++ // add one to round up
+	}
+	// todo: stop rounding up. Here we're rounding up because the calculation for
+	// tx bytes isn't perfect. This catches those edge cases where we
+	// estimate the exact number of shares in the square, when in reality we're
+	// one byte over the number of shares in the square size. This will also cause
+	// blocks that are one square size too big instead of being perfectly snug.
+	// The estimation must be perfect or greater than what the square actually
+	// ends up being.
+	if txShares > 0 {
+		txShares++
+	}
+
+	for _, e := range evd.Evidence {
+		evdBytes += e.Size() + shares.DelimLen(uint64(e.Size()))
+	}
+
+	evdShares = evdBytes / appconsts.CompactShareContentSize
+	if evdBytes > 0 {
+		evdShares++ // add one to round up
+	}
+
+	// sort the msgSummaries by namespace to order them properly. This is okay to do here
+	// as we aren't sorting the actual txs, just their summaries for more
+	// accurate estimations
+	sort.Slice(msgSummaries, func(i, j int) bool {
+		return bytes.Compare(msgSummaries[i].namespace, msgSummaries[j].namespace) < 0
+	})
+
+	// isolate the sizes as we no longer need the namespaces
+	msgShares := make([]int, len(msgSummaries))
+	for i, summary := range msgSummaries {
+		msgShares[i] = summary.size
+	}
+	return txShares, evdShares, msgShares
+}
+
+// overEstimateMalleatedTxSize estimates the size of a malleated tx. The formula it uses will always over estimate.
+func overEstimateMalleatedTxSize(txLen, msgLen, sharesCommitments int) int {
+	// the malleated tx uses the original txLen to account for meta data from
+	// the original tx, but removes the message and extra share commitments that
+	// are in the wire message by subtracting msgLen and all extra share
+	// commitments.
+	malleatedTxLen := txLen - msgLen - ((sharesCommitments - 1) * appconsts.ShareCommitmentBytes)
+	// we need to ensure that the returned number is at least larger than or
+	// equal to the actual number, which is difficult to calculate without
+	// actually malleating the tx
+	return appconsts.MalleatedTxBytes + appconsts.MalleatedTxEstimateBuffer + malleatedTxLen
+}
+
+func nextPowerOfTwo(v int) int {
+	k := 1
+	for k < v {
+		k = k << 1
+	}
+	return k
+}