zips/zip-0401.rst

::

  ZIP: 401
  Title: Addressing Mempool Denial-of-Service
  Owners: Daira Hopwood <daira@electriccoin.co>
  Status: Final
  Category: Network
  Created: 2019-09-09
  License: MIT


Terminology
===========

The key words "MUST", "SHOULD", and "MAY" in this document are to be interpreted
as described in RFC 2119. [#RFC2119]_


Abstract
========

This proposal specifies a change to the behaviour of ``zcashd`` nodes intended to
mitigate denial-of-service from transaction flooding.


Motivation
==========

Adoption of this proposal would increase robustness of Zcash nodes against
denial-of-service attack, in particular attacks that attempt to exhaust node
memory.

Bitcoin Core added size limitation for the mempool in version 0.12
[#BitcoinCore-PR6722]_, defaulting to 300 MB. This was after Zcash forked from
Bitcoin Core.


Requirements
============

The memory usage of a node’s mempool should be bounded.

The eviction policy should as far as possible not be “gameable” by an adversary,
i.e. an adversary should not be able to cause legitimate transactions (that do not
themselves present any denial-of-service problem) to be preferentially evicted
relative to its own transactions.

Any configuration options should have reasonable defaults, i.e. without changing
relevant configuration, a node should be adequately protected from denial-of-service
via mempool memory exhaustion.


Non-requirements
================

The current architecture of Zcash imposes fundamental limits on scaling of
transaction throughput. This proposal does not increase the aggregate transaction
capacity of the network. (The Blossom upgrade does increase transaction capacity,
by a factor of two [#zip-0208]_.)

Denial-of-service issues in the messaging layer of the peer-to-peer protocol are
out of scope for this proposal.

This proposal is focused primarily on memory exhaustion attacks. It does not
attempt to use fees to make denial-of-service economically prohibitive, since that
is unlikely to be feasible while maintaining low fees for legitimate users. It
does not preclude changes in fee policy.


Specification
=============

This specification describes the intended behaviour of ``zcashd`` nodes. Other node
implementations MAY implement the same or similar behaviour, but this is not a
requirement of the network protocol. Thus, RFC 2119 conformance keywords below are
to be interpreted only as placing requirements on the ``zcashd`` implementation (and
potentially other implementations that have adopted this specification in full).

The mempool of a node holds a set of transactions. Each transaction has a *cost*,
which is an integer defined as:

  max(serialized transaction size in bytes, 4000)

Each transaction also has an *eviction weight*, which is *cost* + *low_fee_penalty*,
where *low_fee_penalty* is 16000 if the transaction pays a fee less than the
conventional fee, otherwise 0. The conventional fee is currently defined as
1000 zatoshis [#zip-0313]_.

Each node also MUST hold a FIFO queue RecentlyEvicted of pairs (txid, time), where
the time indicates when the transaction with the given txid was evicted. The txid
(rather than the wtxid as defined in [#zip-0239]_) is used even for version 5
transactions after activation of NU5 [#zip-0252]_.

The RecentlyEvicted queue SHOULD be empty on node startup. The size of RecentlyEvicted
SHOULD never exceed ``eviction_memory_entries`` entries, which is the constant 40000.

There MUST be a configuration option ``mempooltxcostlimit``, which SHOULD default
to 80000000.

There MUST be a configuration option ``mempoolevictionmemoryminutes``, which
SHOULD default to 60.

On receiving a transaction:

* If it is in RecentlyEvicted, the transaction MUST be dropped.
* Calculate its cost. If the total cost of transactions in the mempool including
  this one would exceed ``mempooltxcostlimit``, then the node MUST repeatedly
  call EvictTransaction (with the new transaction included as a candidate to evict)
  until the total cost does not exceed ``mempooltxcostlimit``.

EvictTransaction MUST do the following:

* Select a random transaction to evict, with probability in direct proportion to
  eviction weight.
* Add the txid and the current time to RecentlyEvicted, dropping the oldest entry
  in RecentlyEvicted if necessary to keep it to at most ``eviction_memory_entries``
  entries.
* Remove it from the mempool.

Nodes SHOULD remove transactions from RecentlyEvicted that were evicted more than
``mempoolevictionmemoryminutes`` minutes ago. This MAY be done periodically,
and/or just before RecentlyEvicted is accessed when receiving a transaction.


Rationale
=========

The accounting for transaction size should include some overhead per transaction,
to reflect the cost to the network of processing them (proof and signature
verification; networking overheads; size of in-memory data structures). The
implication of not including overhead is that a denial-of-service attacker would
be likely to use minimum-size transactions so that more of them would fit in a
block, increasing the unaccounted-for overhead. A possible counterargument would
be that the complexity of accounting for this overhead is unwarranted given that
the format of a transaction already imposes a minimum size. However, the proposed
cost function is almost as simple as using transaction size directly.

The threshold 4000 for the cost function is chosen so that the size in bytes of a
typical fully shielded Sapling transaction (with, say, 2 shielded outputs and up
to 5 shielded inputs) will fall below the threshold. This has the effect of
ensuring that such transactions are not evicted preferentially to typical
transparent transactions because of their size.

The proposed eviction policy differs significantly from that of Bitcoin Core
[#BitcoinCore-PR6722]_, which is primarily fee-based. This reflects differing
philosophies about the motivation for fees and the level of fee that legitimate
users can reasonably be expected to pay. The proposed eviction weight function
does involve a penalty for transactions with a fee lower than the standard
(0.0001 ZEC) value, but since there is no further benefit to increasing the fee
above the standard value, it creates no pressure toward escalating fees. For
transactions up to 4000 bytes, this penalty makes a transaction that pays less
than the standard fee value five times as likely to be chosen for eviction
(because 4000 + 16000 = 20000 = 4000 \* 5).

The fee penalty is not included in the cost that determines whether the mempool
is considered full. This ensures that a DoS attacker does not have an incentive
to pay less than the standard fee in order to cause the mempool to be considered
full sooner.

The default value of 80000000 for ``mempooltxcostlimit`` represents no more
than 40 blocks’ worth of transactions in the worst case, which is the default
expiration height after the Blossom network upgrade [#zip-0208]_. It would serve
no purpose to make it larger.

The ``mempooltxcostlimit`` is a per-node configurable parameter in order to
provide flexibility for node operators to change it either in response to
attempted denial-of-service attacks, or if needed to handle spikes in transaction
demand. It may also be useful for nodes running in memory-constrained environments
to reduce this parameter.

The limit of ``eviction_memory_entries`` = 40000 entries in RecentlyEvicted bounds
the memory needed for this data structure. Since a txid is 32 bytes and a
timestamp 8 bytes, 40000 entries can be stored in ~1.6 MB, which is small compared
to other node memory usage (in particular, small compared to the maximum memory
usage of the mempool itself under the default ``mempooltxcostlimit``).
``eviction_memory_entries`` entries should be sufficient to mitigate any
performance loss caused by re-accepting transactions that were previously evicted.
In particular, since a transaction has a minimum cost of 4000, and the default
``mempooltxcostlimit`` is 80000000, at most 20000 transactions can be in the
mempool of a node using the default parameters. While the number of transactions
“in flight” or across the mempools of all nodes in the network could exceed this
number, we believe that is unlikely to be a problem in practice.

Note that the RecentlyEvicted queue is intended as a performance optimization
under certain conditions, rather than as a DoS-mitigation measure in itself.

The default expiry of 40 blocks after Blossom activation represents an expected
time of 50 minutes. Therefore (even if some blocks are slow), most legitimate
transactions are expected to expire within 60 minutes. Note however that an
attacker’s transactions cannot be relied on to expire.


Deployment
==========

This specification is proposed to be implemented in ``zcashd`` v2.1.0. It is
independent of the Blossom network upgrade.


Reference implementation
========================

* `PR 4145: Implementation <https://github.com/zcash/zcash/pull/4145>`_
* `PR 4166: macOS compliation fix <https://github.com/zcash/zcash/pull/4166>`_


References
==========

.. [#RFC2119] `RFC 2119: Key words for use in RFCs to Indicate Requirement Levels <https://www.rfc-editor.org/rfc/rfc2119.html>`_
.. [#zip-0208] `ZIP 208: Shorter Block Target Spacing <zip-0208.rst>`_
.. [#zip-0239] `ZIP 239: Relay of Version 5 Transactions <zip-0239.rst>`_
.. [#zip-0252] `ZIP 252: Deployment of the NU5 Network Upgrade <zip-0252.rst>`_
.. [#zip-0313] `ZIP 313: Reduce Conventional Transaction Fee to 1000 zatoshis <zip-0313.rst>`_
.. [#BitcoinCore-PR6722] `Bitcoin Core PR 6722: Limit mempool by throwing away the cheapest txn and setting min relay fee to it <https://github.com/bitcoin/bitcoin/pull/6722>`_