Accepted
Tendermint Core has a reactor and data structure, mempool, that facilitates the
ephemeral storage of uncommitted transactions. Honest nodes participating in a
Tendermint network gossip these uncommitted transactions to each other if they
pass the application's CheckTx
. In addition, block proposers select from the
mempool a subset of uncommitted transactions to include in the next block.
Currently, the mempool in Tendermint Core is designed as a FIFO queue. In other words, transactions are included in blocks as they are received by a node. There currently is no explicit and prioritized ordering of these uncommitted transactions. This presents a few technical and UX challenges for operators and applications.
Namely, validators are not able to prioritize transactions by their fees or any incentive aligned mechanism. In addition, the lack of prioritization also leads to cascading effects in terms of DoS and various attack vectors on networks, e.g. cosmos/cosmos-sdk#8224.
Thus, Tendermint Core needs the ability for an application and its users to prioritize transactions in a flexible and performant manner. Specifically, we're aiming to either improve, maintain or add the following properties in the Tendermint mempool:
Note, not all of these properties will be addressed by the proposed changes in this ADR. However, this proposal will ensure that any unaddressed properties can be addressed in an easy and extensible manner in the future.
At the core of the v0
mempool reactor is a concurrent linked-list. This is the
primary data structure that contains Tx
objects that have passed CheckTx
.
When a node receives a transaction from another peer, it executes CheckTx
, which
obtains a read-lock on the *CListMempool
. If the transaction passes CheckTx
locally on the node, it is added to the *CList
by obtaining a write-lock. It
is also added to the cache
and txsMap
, both of which obtain their own respective
write-locks and map a reference from the transaction hash to the Tx
itself.
Transactions are continuously gossiped to peers whenever a new transaction is added
to a local node's *CList
, where the node at the front of the *CList
is selected.
Another transaction will not be gossiped until the *CList
notifies the reader
that there are more transactions to gossip.
When a proposer attempts to propose a block, they will execute ReapMaxBytesMaxGas
on the reactor's *CListMempool
. This call obtains a read-lock on the *CListMempool
and selects as many transactions as possible starting from the front of the *CList
moving to the back of the list.
When a block is finally committed, a caller invokes Update
on the reactor's
*CListMempool
with all the selected transactions. Note, the caller must also
explicitly obtain a write-lock on the reactor's *CListMempool
. This call
will remove all the supplied transactions from the txsMap
and the *CList
, both
of which obtain their own respective write-locks. In addition, the transaction
may also be removed from the cache
which obtains it's own write-lock.
When considering which approach to take for a priority-based flexible and
performant mempool, there are two core candidates. The first candidate is less
invasive in the required set of protocol and implementation changes, which
simply extends the existing CheckTx
ABCI method. The second candidate essentially
involves the introduction of new ABCI method(s) and would require a higher degree
of complexity in protocol and implementation changes, some of which may either
overlap or conflict with the upcoming introduction of ABCI++.
For more information on the various approaches and proposals, please see the mempool discussion.
The Ethereum mempool, specifically Geth,
contains a mempool, *TxPool
, that contains various mappings indexed by account,
such as a pending
which contains all processable transactions for accounts
prioritized by nonce. It also contains a queue
which is the exact same mapping
except it contains not currently processable transactions. The mempool also
contains a priced
index of type *txPricedList
that is a priority queue based
on transaction price.
The Diem mempool
contains a similar approach to the one we propose. Specifically, the Diem mempool
contains a mapping from Account:[]Tx
. On top of this primary mapping from account
to a list of transactions, are various indexes used to perform certain actions.
The main index, PriorityIndex
. is an ordered queue of transactions that are
“consensus-ready” (i.e., they have a sequence number which is sequential to the
current sequence number for the account). This queue is ordered by gas price so
that if a client is willing to pay more (than other clients) per unit of
execution, then they can enter consensus earlier.
To incorporate a priority-based flexible and performant mempool in Tendermint Core,
we will introduce new fields, priority
and sender
, into the ResponseCheckTx
type.
We will introduce a new versioned mempool reactor, v1
and assume an implicit
version of the current mempool reactor as v0
. In the new v1
mempool reactor,
we largely keep the functionality the same as v0
except we augment the underlying
data structures. Specifically, we keep a mapping of senders to transaction objects.
On top of this mapping, we index transactions to provide the ability to efficiently
gossip and reap transactions by priority.
We introduce the following new fields into the ResponseCheckTx
type:
message ResponseCheckTx {
uint32 code = 1;
bytes data = 2;
string log = 3; // nondeterministic
string info = 4; // nondeterministic
int64 gas_wanted = 5 [json_name = "gas_wanted"];
int64 gas_used = 6 [json_name = "gas_used"];
repeated Event events = 7 [(gogoproto.nullable) = false, (gogoproto.jsontag) = "events,omitempty"];
string codespace = 8;
+ int64 priority = 9;
+ string sender = 10;
}
It is entirely up the application in determining how these fields are populated
and with what values, e.g. the sender
could be the signer and fee payer
of the transaction, the priority
could be the cumulative sum of the fee(s).
Only sender
is required, while priority
can be omitted which would result in
using the default value of zero.
The existing concurrent-safe linked-list will be replaced by a thread-safe map
of <sender:*Tx>
, i.e a mapping from sender
to a single *Tx
object, where
each *Tx
is the next valid and processable transaction from the given sender
.
On top of this mapping, we index all transactions by priority using a thread-safe
priority queue, i.e. a max heap.
When a proposer is ready to select transactions for the next block proposal,
transactions are selected from this priority index by highest priority order.
When a transaction is selected and reaped, it is removed from this index and
from the <sender:*Tx>
mapping.
We define Tx
as the following data structure:
type Tx struct {
// Tx represents the raw binary transaction data.
Tx []byte
// Priority defines the transaction's priority as specified by the application
// in the ResponseCheckTx response.
Priority int64
// Sender defines the transaction's sender as specified by the application in
// the ResponseCheckTx response.
Sender string
// Index defines the current index in the priority queue index. Note, if
// multiple Tx indexes are needed, this field will be removed and each Tx
// index will have its own wrapped Tx type.
Index int
}
Upon successfully executing CheckTx
for a new Tx
and the mempool is currently
full, we must check if there exists a Tx
of lower priority that can be evicted
to make room for the new Tx
with higher priority and with sufficient size
capacity left.
If such a Tx
exists, we find it by obtaining a read lock and sorting the
priority queue index. Once sorted, we find the first Tx
with lower priority and
size such that the new Tx
would fit within the mempool's size limit. We then
remove this Tx
from the priority queue index as well as the <sender:*Tx>
mapping.
This will require additional O(n)
space and O(n*log(n))
runtime complexity. Note that the space complexity does not depend on the size of the tx.
We keep the existing thread-safe linked list as an additional index. Using this index, we can efficiently gossip transactions in the same manner as they are gossiped now (FIFO).
Gossiping transactions will not require locking any other indexes.
Performance should largely remain unaffected apart from the space overhead of keeping an additional priority queue index and the case where we need to evict transactions from the priority queue index. There should be no reads which block writes on any index
There are a few considerable ways in which the proposed design can be improved or
expanded upon. Namely, transaction gossiping and for the ability to support
multiple transactions from the same sender
.
With regards to transaction gossiping, we need empirically validate whether we need to gossip by priority. In addition, the current method of gossiping may not be the most efficient. Specifically, broadcasting all the transactions a node has in it's mempool to it's peers. Rather, we should explore for the ability to gossip transactions on a request/response basis similar to Ethereum and other protocols. Not only does this reduce bandwidth and complexity, but also allows for us to explore gossiping by priority or other dimensions more efficiently.
Allowing for multiple transactions from the same sender
is important and will
most likely be a needed feature in the future development of the mempool, but for
now it suffices to have the preliminary design agreed upon. Having the ability
to support multiple transactions per sender
will require careful thought with
regards to the interplay of the corresponding ABCI application. Regardless, the
proposed design should allow for adaptations to support this feature in a
non-contentious and backwards compatible manner.
ResponseCheckTx
Protocol Buffer type.CheckTx
but not end up being committed in a block because they fail
CheckTx
later. e.g. Consider Tx1 that sends funds from existing
account Alice to a new account Bob with priority P1 and then later
Bob's new account sends funds back to Alice in Tx2 with P2,
such that P2 > P1. If executed in this order, both
transactions will pass CheckTx
. However, when a proposer is ready to select
transactions for the next block proposal, they will select Tx2 before
Tx1 and thus Tx2 will fail because Tx1 must
be executed first. This is because there is a causal ordering,
Tx1 ➝ Tx2. These types of situations should be rare as
most transactions are not causally ordered and can be circumvented by simply
trying again at a later point in time or by ensuring the "child" priority is
lower than the "parent" priority. In other words, if parents always have
priories that are higher than their children, then the new mempool design will
maintain causal ordering.CheckTx
and entered the mempool can later be evicted
at a future point in time if a higher priority transaction entered while the
mempool was full.