Solana has announced three main changes in its mitigation plan to address the stability and resilience of the network:
The measures are targeting the intense traffic responsible for two out of the three recent incidents. Although the changes being proposed by Solana developers are considered abstract or deeply technical for the general part of the community, the concepts are not completely new, being imported from other already mature systems. In this article, we will try to break down the technicalities and explain them in simple terms.
The current Solana client version for validator nodes (v1.10) already paves the way for some of these improvements to be iterated on until optimal market fit. Fee prioritization is targeted for the v1.11 release, according to the official announcement.
Solana used to adopt the User Datagram Protocol (UDP) for transmitting transactions between nodes in the network. Nodes send transactions through UDP directly to the leader — the staked node responsible for proposing the block in that particular slot — without a previous connection being established. UDP does not handle traffic congestion or delivery confirmation for data. In situations of network congestion, the leader is unable to handle the volume of incoming traffic, which means some packets get dropped. Even at quiet times, some level of packet loss is normal. By sending the same transaction multiple times, users have a greater chance that at least one of their attempts will arrive.
In contrast to UDP is the Transmission Control Protocol (TCP). TCP includes more sophisticated features but for this to work, it requires a session (i.e. a known connection was previously established between the client and the server). The receiver acknowledges (“acks”) packets and the sender knows when to stop sending packets in case of intense traffic. TCP allows for re-transmitting lost packets, once the sender stops receiving acks, the interpretation is that something must be lost, so the sender should slow down.
TCP is not ideal for some use cases though. In particular, it sequences all traffic. If one portion of the data is lost, everything after it needs to wait. That is not great for Solana transactions, which are independent.
QUIC is a general-purpose protocol which is used by more than half of all connections from the Chrome web browser to Google’s servers. QUIC is the name of the protocol, not an acronym.
QUIC is an alternative to TCP with similar features: a session, which then enables backpressure to slow the sender down, but it also has a concept of separate streams; so if one transaction gets dropped, it doesn’t need to block the remaining ones.
Solana is a permissionless network. Anyone running a Solana client is a “node” in the network and is able to send messages to the leader. Nodes can operate as validators — when it is signing and sending votes — and (or) they can expose their RPC (Remote Procedure Call) interface to receive messages from applications such as wallets and DEXs, and send those to the leader.
The leader listens on a UDP port and RPCs listen on a TCP port. Given the leader schedule is public, sophisticated players with algorithmic strategies (“bots”) are able to send transactions to the leader directly, bypassing any additional RPC nodes that would only increase latency. With the leader being spammed, the network gets congested and that deteriorates performance. The UDP port used by the leader will be replaced by a QUIC port.
Quality of Service (“QoS”) is the practice of prioritizing certain types of traffic when there is more traffic than the network can handle.
Last January, after Solana faced performance issues as automated trading strategies (aka “liquidator bots”) spammed the network with more than 2 million packets per second, mostly duplicate messages, Anatoly Yakovenko mentioned in a tweet that they would bring the QoS concept to Solana.
The Leader currently tries to process transactions as soon as they arrive. Because IPs are verifiable through QUIC, validators will be able to prioritize and limit the traffic for specific connections. Instead of validators and RPCs blasting transactions at the leader as fast as they can, effectively DoS’ing the leader, they would have a persistent QUIC connection. If the network (IP) gets congested, it will be possible to identify and apply policies to large traffic connections, limiting the number of messages the node can send (“throttle”). These policies are known as QoS.
Internally, staked weighted QoS means queuing transactions in different channels depending on the sender, weighted by the amount of SOL staked. Non-staked nodes will then be incentivized to send transactions to staked nodes first, instead of sending directly to the leader, for a better chance of finding execution, since excess messages from non-staked nodes will most likely be dropped by the leader.
According to Anatoly validators will be responsible for shaping their own traffic, and applying policies that will avoid vulnerability. For example, if a particular node sends huge amounts of transactions, even if they are staked, validators can take action, ignoring the connections established with this node in order to protect network performance.
Solana fees are currently fixed and charged for each signature required in a transaction (5000 lamports = 0.000005 SOL). If there is high competition in a specific market, users face the risk of not getting transactions executed. With a fixed transaction fee, there is no way to communicate priority or compete by paying more to get their transaction prioritized. Without alternatives, users (usually bots) spam transactions to the leader (and soon-to-be leaders) in hope that at least one of them is successful. In many situations, this behavior generates more traffic than what the network can process.
A priority fee is soon to be included in Solana, allowing users to specify an arbitrary “additional fee” to be collected upon execution of the transaction and its inclusion in a block. This mechanism would not only help the network to prioritize time-sensitive transactions but also tends to reduce the amount of invalid or duplicated messages sent by algorithms since speculative operations can become unprofitable with an increase in the total cost.
The ratio of this fee to the requested compute units (the computational cost to the program to perform all operations) will serve as a transaction’s execution priority weight. This ratio will be used by nodes to prioritize the transactions they send to the leader. Additional fees will be treated identically to the base fee today: 50% of the fees paid will be collected by the leader and 50% will be burned.
At this point, you could think of several blocks being filled only with transactions targeting an NFT mint. However, there is a limit time for each account to be locked for writing on a single slot (600 to 800 milliseconds). The remnant block space can be filled with transactions writing in different accounts, even if they offer a smaller priority fee. High-priority transactions trying to write to an account that has already reached its limit will be included in the next block.
Each Solana transaction specifies the writable accounts — the portion of the state that will be modified. This allows transactions to be executed in parallel, as long as transactions are independent, i.e. do not access the same accounts. If two transactions write or read to the same account, these two transactions can not be processed in parallel, because they affect the same state.
The Solana team argues that the priority fee will then behave as parallel auctions, affecting only the “hot market” instead of the global price, allowing the fee to grow for a specific queue of transactions trying to write in that account only.
How does the user know the fee to adopt to get a mint? RPCs nodes will need to estimate an adequate fee, most likely using a simple statistical method, for example averaging the actual cost of similar transactions in previous N blocks, or even a quantile. The optimal method will depend on the market, and whether fees for similar transactions are more volatile (high demand) or stable (less demand).
In practice, the priority fee can have a global effect, if the parallel auctions are not implemented on the validator client. With RPCs and users being responsible for arbitrarily setting it, during high intense traffic, applications will likely try to get priority even though they do not interact with the “hot market”, causing an increase in the fee price for other lower demand dApps.
Fee prioritization is targeted for the v1.11 release, according to the official announcement.
The present article covered the three pieces Solana is actively working on to deal with congestion issues, which include changing the communication protocol from UDP to QUIC, adding stake-weighted QoS for transaction prioritization and a fee market that increases fees with high demand. All of these 3 improvements aspire to improve the performance of Solana, which has been experiencing degraded performance quite often.
We hope it was possible to clarify these concepts and understand the motivations and choices being made. Exploring Solana source code would be an essential next step to investigate the exact metrics being implemented in QoS to select or drop transactions or the mechanism behind the increase (and decrease) of fees and other questions that remain unanswered.
I would like to thank the Chorus One team for the enlightening discussions and knowledge sharing, especially Ruud van Asseldonk for the technical review, and Xavier Meegan for the support.
This is the second article of the Solana MEV outlook series. In this series, we use a subset of transactions to extrapolate which type of Maximum Extractable Value (MEV) is being extracted on the Solana network and by whom.
MEV is an extensive field of research, ranging from opportunities created by network design or application-specific behaviour to trading strategies similar to those applied in the traditional financial markets. As a starting point, our attempt is to investigate if sandwich attacks are happening. In the first article, we examined Orca’s swap transactions searching for evidence of this pattern. Head to Solana MEV Outlook — part 1 for a detailed introduction, goals, challenges and methodology. A similar study is performed in the present article. We are going to look at on-chain data, considering approximately 8 h of transactions on the Raydium DEX. Given the magnitude of 4 x 10⁷ transactions per day, considering only Decentralized Exchanges (DEX) applications on the Solana ecosystem. This simplification is done to get familiarity with data, extrapolating as much information as we can to extend towards a future analysis by employing a wider range of transactions.
Raydium is a relevant Automated Market Maker (AMM) application on the Solana ecosystem, the second program in the number of daily active users and the third in terms of program activity.
Raydium program offers two different swap instructions:
Although the user interface (“UI”) interacting with the smart contract sets the swap instruction to use the first instruction type, leaving SwapBaseIn responsible for 99.9% of successfully executed swap instructions:
We built a dataset, extracting the inputs from the data byte array passed to the program, and the actual swap token amounts by looking at the instructions contained in the transaction. Comparing the minimum amount of tokens specified in the transaction and the actual amount the user received, we estimate the maximum slippage tolerance for every transaction. By computing the corresponding slippage, we obtain the histogram:
The default value for slippage on the Raydium App is set to 1%. We can assume that at least 28% of transactions use the default value. Since it is not possible to know the state of the pool when creating the transaction, this number could be a bit higher.
It can be assumed that nearly 0% of slippage values are only achieved by sophisticated investors using automated trading strategies. Orca swaps’ histogram, presented in Fig 2.2 of the previous article, shows a peak in transactions with slippage of around 0.1%. On Raydium, a relevant proportion of transactions lies below 0.05%. This fact can suggest that trading strategies with lower risk tolerance, i.e price-sensitive strategies correspond to 25% of the swaps transactions (accumulating the first two bars in the histogram).
Other evidence of automated trading being common on this DEX is that on average, 40% of transactions fail, mostly because of the tight slippage allowed by user settings.
We are considering more than 30,000 instructions interacting with the Raydium AMM program, from time 02:43:41 to time 10:25:21 of 2022–04–06 UTC. For statistics purposes, failed transactions are ignored.
Although 114 different liquidity pools are accessed during this period, the SOL/USDC pool is the most traded pool, with 4,000 transactions.
The sample contains 1366 different validators as leaders in more than 35000 slots we are considering, representing 93% of the total stake and 78% of the total validator population by the time of writing, according to Solana Beach.
Of 5,101 different addresses executing transactions, 10 accounts concentrate 23% of the total transactions. One of the most active accounts on Raydium, Cwy…3tf also appears in the top 5 accounts in Orca DEX.
The graph below shows the total number of transactions for accounts with at least two transactions in the same slot. If used as a proxy to identify automated trading, on average 9 different accounts can be classified:
We can also look at the pools where these accounts execute more often. It is possible to notice they tend to specialize in different pools. The table below shows the two pools with more transactions for each of the 5 more active addresses:
By deep-diving into account activity by pool, we can see that two accounts concentrate transactions on WSOL/USDT pool; one account is responsible for half of all transactions in the mSOL/USDC pool; most of the transactions in the GENE/RAY pool are done by only one account (Cwy…3tf).
Searching for sandwich behaviour means we need to identify at least 3 transactions executed in the same pool in a short period of time. For the purpose of this study, only consecutive transactions would be considered. The strategy implies the first transaction to be in the same direction of the sandwiched transaction and a transaction in the opposite direction of the initial trade, closing out the positions of the MEV player.
The need for price impact implies a dependence on the amount of capital available to be used in every trade. Some MEV strategies can be performed atomically, with a sequence of operations executed in the same transaction. These strategies usually benefit from flash loans, allowing for anyone to apply it disregarding the capital they have access to. This is not the case for sandwich attacks, since the profit is realized after the successful execution of all the transactions (Fig. 10).
As shown in the first article, the amount of capital needed in order to create value depends on the Total Value Locked in the pool — the deeper the liquidity, the more difficult it is to impact the price. Head to Fig. 2.4 of the first article for the results of simulation into the Orca’s SOL/USDC pool. The figure shows the initial capital needed in order to extract a given percentage of the swap.
In the current sample, we have found 129 blocks with more than three swaps in the same pool, most of the swaps are in the same direction — no evidence of profit-taking. As shown in Fig. 11 below, the pool SAMO_RAY is the pool with more occurrences of multiple swaps in the same slot.
When searching for blocks and pools with swaps in opposite directions as a proxy to profit-taking, 9 occurrences are left with a potential sandwich attack pattern, as shown in the table below (Fig 12). After further investigation of the transactions and the context in which the instructions were executed, it is fair to assume the operations are related to arbitrage techniques between different trading venues or pools.
In this report, we were able to access the activity of the Raydium DEX. The conclusions are based on a limited amount of data, assuming our sample is comprehensive enough to reflect the general practices involving the dApp.
It is possible to notice relevant activity from automated trading and price-sensitive strategies such as arbitrage, which corresponds to 25% of swap transactions. On average, only 40% of transactions are successfully executed and 72% of all reverted transactions fail because of small slippage tolerance. Approximately, 28% of transactions can be classified as manual trading, since they use the default slippage value.
Of 5101 different accounts interacting with the Raydium program, 10 accounts concentrate 23% of the total transactions. One of the most active accounts on Raydium, Cwy…3tf also appears in the top 5 accounts in Orca DEX transactions. This same account is responsible for 77% of swaps in the GENE/RAY pool.
There were 9 occurrences of a potential pattern of a Sandwich attack discarded after further investigation.
It is important to mention that this behaviour is not only dependent on the theoretical possibility but largely biased by market conditions. The results in $13m MEV during Wormhole Incident and $43m Total MEV from Luna/ UST Collapse on Solana demonstrate the increase in profit extracted from MEV opportunities during stressful scenarios. Although the study focuses attention on different strategies and does not mention sandwich attacks, the probability of this strategy happening can also increase, given the smaller liquidity in pools (TVL) and the occurrence of trades with bigger size and slippage tolerance.
This is my first published article. I hope you enjoyed it. If you have questions, leave your comment below and I will be happy to help.
Solana is a young blockchain, and having a complete picture of what is happening on-chain is a difficult task — especially due to the high number of transactions daily processed. The current number of TPS is around 2,000, meaning that we need to deal with ~ 10⁸ transactions per day, see Fig. 1.1.
When processing transactions, we have to deal with the impossibility of a-priori knowing its status before querying information from an RPC node. This means that we are forced to process both successful and failed transactions. The failed transactions, most of which come from spamming bots that are trying to make a profit (e.g. NTF, arbitrage, etc.), constitutes ~ 20% of the successful ones. The situation slightly improves if we consider only program activity. By only considering what happens on Decentralized Exchanges (DEXs), we are talking about 4x10⁷ transactions per day, see Fig. 1.2. This makes it clear that a big effort is required to assess which type of Maximum Extractable Value (MEV) attack is taking place and who is taking advantage of it, even because tools like Flashbots do not exist on Solana.
In what follows, we are going to estimate what happened on-chain considering only ~5 h of transactions on Orca DEX, from 11:31:41 to 16:34:19 on 2022–03–14. This simplification is done to get familiarity with data, extrapolating as much information as we can to extend towards a future analysis by employing a wider range of transactions. It is worth mentioning that Orca DEX is not the program with the highest number of processed instructions, which indicates that a more careful analysis is needed to look also into other DEX — this is left for future study.
The aim of this preliminary analysis is to gain familiarity with the information contained in usual swap transactions. One of our first attempts is to extrapolate if sandwich attacks are happening, and if so, with which frequency. In Section 2, we are going to look at the anatomy of a swap transaction, focussing on the type of sandwich swap in section 2.1. Section 2.2 is devoted to the description of “actors” that can make a sandwich attack. In Section 3, we describe the dataset employed, leaving the description of the results in Section 4. Conclusions are drawn in Section 5.
On Solana, transactions are made by one or more instructions. Each instruction specifies the program that executes them, the accounts involved in the transaction, and a data byte array that is passed to the program. It is the program’s task to interpret the data array and operate on the accounts specified by the instructions. Once a program starts to operate, it can return only two possible outcomes: success or failure. It is worth noticing that an error return causes the entire transaction to fail immediately. For more details about the general anatomy of the transaction see the Solana documentation.
To decode each of the instructions we need to know how the specific program is written. We know that Orca is a Token Swap Program, thus we have all the ingredients needed to process data. Precisely, taking a look at the token swap instruction, we can immediately see that a generic swap takes as input the amount of token that the user wants to swap, and the minimum amount of token in output needed to avoid excessive slippage, see Fig. 2.1.
The minimum amount of tokens in output is related to the actual number of tokens in output by the slippage S, i.e.
from which
Thus, we can extract the token in input and the minimum token in output from the data byte array passed to the program, and the actual token in output by looking at the instructions contained in the transaction.
By computing the corresponding slippage defined in Eq. (2.2) we obtain the histogram in Fig. 2.2. From this picture, we can extrapolate different information. The first one is, without doubt, the distribution of transactions around the default value of slippage on Orca, i.e. 0.1%, 0.5% and 1%. This makes complete sense since the “common-user” is prone to use default values, without spending time in customization. The second one is the preference of users to select the lowest value for the slippage. The last one concerns the shape of the tails around the default values. A more detailed analysis is needed here since it is not an easy task to have access to what actually is contained inside them. The shape surely depends on the bid/ask scatter, which is a pure consequence of the market dynamic. The tails may also contain users that select a different slippage with respect to the default values. However, one thing is assured: this histogram contains swaps from which the slippage can yet be extracted. As we will see, from this we can extrapolate an estimate of the annualized revenue due to sandwich attacks.
The goal of this report is to search for hints of sandwich swaps happening on Orca DEX. All findings will be used for future research, thus we think it is useful to define what we refer to as sandwich swaps and how can someone take advantage of them.
Let’s start with its basic definition. Let’s assume a user (let’s say Alice) wants to buy a token X on a DEX that uses an automated market maker (AMM) model. Let’s now assume that an adversary sees Alice’s transaction (let’s say Bob) and can create two of its own transactions which it inserts before and after Alice’s transaction (sandwiching it). In this configuration, Bob buys the same token X, which pushes up the price for Alice’s transaction, and then the third transaction is the adversary’s transaction to sell token X (now at a higher price) at a profit, see Fig. 2.3. This mechanism works until the price at which Alice buys X remain sbelow the value X･(1+S), where S represents the slippage set by Alice when she sends the swap transaction to the DEX.
Since Bob needs to increase the value of the token X inside the pool where Alice is performing the swap, it is evident that the core swaps inserted by Bob should live on the same pool employed by Alice.
From the example above, it may happen that Bob does not have the capital needed to significantly change the price of X inside the pool. Suppose that the pool under scrutiny regards the pair X/Y and that the AMM implements a constant product curve. In the math formula we have:
where k is the curve invariant. If we set the number of tokens Y in the pool equal to 1,000,000 and the number of tokens X equal to 5,000,000 and assuming that Alice wants to swap 1,000 token Y, we have that the amount of token X in output is:
It is worth noting that here we are not considering the fee that is usually paid by the user. If Alice set a slippage of 5%, this means that the transaction will be executed until the output remains above 4'745.25. This means if Bob is trying to take this 5%, he will need an initial capital of 26,000 token Y.
Sometimes this capital may be inaccessible, allowing Bob to only take a portion of the 5% slippage. For example, let’s consider the Orca pool SOL/USDC, with a total value locked (TVL) of $108,982,050.84 at the time of writing. This pool implements a constant product curve, which allows us to use Eqs. (2.3) and (2.4) to simulate a sandwich attack. Fig. 2.4 shows the result of this calculation.
It is clear that the initial capital to invest may not be accessible to everyone. Further, it is important to clarify that the result is swap-amount independent. Indeed, for each amount swapped by Alice, the swap made by Bob is the one that “moves” the prices of the initial tokens inside the pool. The scenario is instead TVL dependent. If we repeat the same simulation for the Orca pool ETH/USDC, with a TVL of $2,765,189.76, the initial capital needed to extract a higher percentage of the slippage of Alice drastically decreases, see Fig. 2.5.
From the example above, let’s consider the case in which Bob has an initial capital of 2,000 token Y. If he is able to buy the token Y before Alice’s transaction, Alice will obtain an output of 4,975.09 token X, which is only 0.4% lower than the original amount defined in Eq. (2.4).
At this point, Bob has another possibility. He can try to order transactions that are buying the same token X after its transaction, but immediately before Alice’s swap. In this way, he can use the capital of other users to take advantage of Alice’s slippage, even if Bob’s initial capital is not enough to do so, see Fig. 2.6. This of course results in a more elaborate attack, but likely to happen if Bob has access to the order book.
It is not an easy task to spot the actors behind a sandwich attack on Solana. In principle, the only profitable attackers are the leaders. This is because there isn’t a mempool, and the only ones that know the exact details of the transactions are the validators that are in charge of writing a block. In this case, it may be easier to spot hints of a sandwich attack. Indeed, if a leader orders the swap transactions to perform a sandwich, it should include all of them in the same block to prevent an unsuccessful sandwich.
The immediately following suspect is the RPC service that the DAPP is using. This is because the RPC service is the first to receive the transaction over HTTP, since it is its role to look up the current leader’s info using the leader schedule and send it to the leader’s Transaction Processing Unit (TPU). In this case, it would be much more difficult to spot hints of sandwiching happening since in principle the swap transactions involved can be far from each other. The only hook we can use to catch the culprit is to spot surrounding transactions made by the same user, which will be related to the RPC. This is a consequence of the lower price fee on Solana, which raises the likelihood that a sandwich attack can happen by chance spamming transactions in a specific pool. This last one is clearly the riskiest since there is no certainty that the sequence of transactions is included in the exact order in which the attacker originally planned it.
Before entering the details of the analysis, it is worth mentioning that, standing on what is reported on Solana Beach, we have a total of 1,696 active validators. Our sample contains 922 of them, i.e. 54.37% of the total validator population. The table below shows the validator that appears as the leader in the time window we are considering. Given the likelihood-by-stake for a validator to be selected as a leader, we retain fair to assume that our sample is a good representation of what’s happening on Orca. Indeed, if a validator is running a modified version of the vote account program to perform sandwich swap, the rate of its success will be related to the amount of staked tokens, not only by actual MEV opportunities. Further, modifying the validator is not an easy task, thus smaller validators will not have the resources to do that. Since we have all the 21 validators with a supermajority plus a good portion of the others (i.e. we are considering half of the current number of active validators), if such a validator exists, its behaviour is easily spotted in our sample. However, it is worth mentioning that a complete overview of the network requires the scrutiny of all validators, without making assumptions of that kind. Such achievement is behind the scope of this report, which aims primarily to explore which type of sandwich can be done and how to spot them.
Having clarified this aspect, we firstly classify the types of swaps that are performed on the Orca DEX. The table below shows the accounts that are performing more than two transactions. It is immediately visible that most of the transactions are done by only 2 accounts over 78 involved.
As explained in Section 1, we are considering 5H of transactions on Orca DEX, from 11:31:41 to 16:34:19 on 2022–03–14. This sample contains a total of 12,106 swaps, with pool distribution in Fig. 3.1.
By deep-diving into the swap, we can see that most of the transactions in the 1SOL/SOL [aq] and 1SOL/USDC [aq] are done by only two accounts, see Fig. 3.2. Here [aq] stands for Aquafarm, i.e. an Orca’s yield farming program. We can also see the presence of some aggregate swaps in the SOL/USDC [aq] and ORCA/USDC [aq] pools.
We started searching for the presence of leaders performing sandwich swaps. As we described in Section 2.1, in general, a swap can happen in two ways. For both of them, if such a type of surrounding is done by a leader, we should see the transactions under scrutiny included in the same block. This is because, if a leader wants to make a profit, the best strategy is to avoid market fluctuations. Further, if the attacker orders the transactions without completing the surrounding, the possibility that another leader reorders transactions cancelling the effect of what was done by the attacker is not negligible.
By looking at the slots containing more than 3 swaps in the same pool, we ended up with 6 slots of that kind, out of 7479. Deep diving into these transactions, we found that there is no trace of a sandwich attack done within the same block (and so, from a specific leader). Indeed, each of the employed transactions is done by a different user, marking no evidence of surrounding swaps done to perform a sandwich attack. The only suspicious series of transactions is included in block # 124899704. We checked that the involved accounts are interacting with the program MEV1HDn99aybER3U3oa9MySSXqoEZNDEQ4miAimTjaW, which seems to be an aggregator for arbitrage opportunities.
As mentioned in Section 2.2, validators are not the only possible actors. Thus, to complete the analysis we also searched for general surrounding transactions, without constraining them to be included in the same block. We find that only 1% of the total swaps are surrounded, but again without strong evidence of actual sandwich attacks (see Fig. 4.1 for the percentage distribution). Indeed, by looking at those transactions it comes out that the amount of token exchanged is too low to be a sandwich attack (see Sec. 2).
Before ending this section, it is worth mentioning that if we extrapolate the annual revenue that a leader obtains by taking 50% of the available slippage for swaps with a slippage greater than 1%, we are talking about an amount of ~ 240,000.00 USD (assuming that the attacker is within the list of 21 validators with supermajority), see Fig. 4.2. Of course, this is not a real estimate since it is an extrapolation from only 5h of transactions, thus we need to stress that the actual revenue can be different. Further, this is not an easily accessible amount due to what we showcased in Sec. 2. However, the amount in revenue clearly paves the way for a new type of protection that validators should offer to users, especially if we take into account that Orca is not the DEX with the highest amount of processed swaps. Since at the moment there is no evidence that swaps are sandwiched, we will take no action in this direction. Instead, we will continue monitoring different DEXs by taking snapshots in different timeframes informing our users if a sandwich attack is spotted on Solana.
In this report, we define two types of sandwich attacks that may happen on a given DEX. We further describe who are the possible actors that can perform such a type of attack on Solana and how to spot them. We analyzed data from ~5 h of transactions on Orca DEX, from 11:31:41 to 16:34:19 on 2022–03–14 (that is, 12,106 swaps). Despite the cutting of the number of transactions employed, we argued why we believe this sample could fairly be a “good” representation of the entire population.
Our findings show no evidence that sandwich attacks are happening on Solana by considering two possibilities. The former is that a validator is running a modified version “trained” to perform a sandwich attack on Orca. The latter is that an RPC is trying to submit surrounding transactions. We discovered that only 1% of transactions are actually surrounded by the same user, but none of them is included in the same block — excluding the possibility that a leader is taking advantage of the slippage. By deep-diving into this, we discover that the amount exchanged by these transactions results are too low for capital to be invested to exploit the slippage and submit a profitable sandwich attack.
We also show how the capital needed to make sandwich attacks profitable may not be accessible to everyone, narrowing the circle of possible actors.
Big thanks to my colleagues at Chorus One for their contributions to this post, especially Umberto Natale for providing a lot of the data used, full report here.
Many different industries are using Ethereum to build new decentralized applications.
2021 was the year when this vision stopped being reserved for a small subset of the population with pre-existing capital (investors) or technical expertise (developers), as the popularity of Ethereum reached new heights.
Artists are disrupting traditional notions of value, with OpenSea (the largest NFT Marketplace) growing its transactions volume by “over 600x” in a year.
People are organizing self-sustainable communities, as DAO members take control over their own financial freedom and digital identity.
Builders are creating never-before-seen decentralized financial assets, where Ethereum-based Uniswap continues to dominate. But real fun also came to crypto gaming: many are playing Dark Forest, a game that experiments with cutting edge scaling technology.
Most recently, the whole crypto community has come together to aid Ukraine, at the same rate as well-established international organizations.
Bringing Ethereum into the sun and serving all of humanity, inevitably requires a scalable, secure, and resilient network.
This post aims to set the stage for Ethereum as it nears its greatest milestone, and to take a peek at the future for what this could mean in impact for both staking providers and delegators, after the Altair upgrade. To do this, we have delved into risks, rewards and the complex network that sits in between.
The big goals for the future of Ethereum are scattered across its official roadmap.
Understanding this design is key to comprehending the associated risks and rewards of PoS Ethereum. The process of upgrading started in December 2020, when the first piece of the puzzle fell into place: the Beacon Chain went live. This PoS system sets the basic consensus mechanism, by assigning the right to create a block through a deterministic lottery process. Staking nodes with a higher balance have more probability to be selected. The rewards for staking include block rewards and transaction fees, and we explore these further in the following section.
To stake in Ethereum and run a validator, 32 ETH needs to be sent to the Ethereum Deposit Contract, along with two key parameters: 1) the validator public key and 2) the withdrawal credentials for the deposit. Critically, the public key and withdrawal credentials do not need to be controlled by the same entity. This allows for two ways to participate in the protocol: as a validator or as a delegator (individuals who pass the responsibility of validation while still earning a portion of the rewards). Staking providers such as Chorus One offer ETH holders the opportunity to stake their tokens and participate in consensus through its platform.
Because chosen stakers are given exclusive rights to create a block, the protocol must consist of measures to counteract malicious attack vectors. The implementation of this consensus mechanism relies on three core elements: A fork-choice rule, a concept of finality, and slashing conditions. It is important to note that in PoS networks, slashing is not a necessary incentive for correct behavior by validators but rather an artifact of the particular block rewards and mechanism implemented. Because rewards are based on blocks processed or accepted, there’s an incentive for a validator to validate all forks in the chain, even conflicting ones (nothing at stake). Therefore, a slashing rule has to be implemented, as a matter of design.
The Altair hard fork of October 2021 introduced additional elements to consensus, namely sync committees. Validators that are part of this committee have the duty of continually signing the block header to allow a new set of light clients to easily sync up at very low computational and data cost. These concepts of head of the chain, target of attestation and source of attestation are critical to finalizing blocks and earning rewards. Checkpoints are set on-chain to achieve these goals: when a checkpoint is finalized, all previous slots are finalized. There is no limit to the number of blocks that can go through this system. A checkpoint can only be finalized after the process of consensus chooses another validator, and the infinite machine starts all over again.
It is likely that you’ve come across the prime assumption for PoS: “validators will be lazy, take bribes, and that they will try to attack the system unless they are otherwise incentivized not to” Hope for the best, but expect the worst.
You may have also seen floating around different figures for the “estimated APR” for running a validator, and wondered — where does this number even come from? All estimations for returns rest on a set of assumptions, and many published calculations were presented using outdated specs for the Beacon Chain.
So, let’s take a current look. Incentives in Altair come in the form of rewards, penalties, and slashings. Of these three, slashing is the most relevant to validator health. While crypto rewards have been around for years, their complexity and adoption have seen a significant rise in the recent past. Offerings do differ platform by platform, and all carry different kinds of risks.
One of the main conceptual reworks of Altair was in redesigning how validators are rewarded (and penalized). The idea was to make these incentives more systematic and simplify state management. But it also ups the ante on validator responsibilities.
At present, validators are rewarded/penalized if they fulfill determined duties:
A validator can propose one attestation and one block at a time. Depending on their properties, the reward varies. Participation in proposals and in the sync committee are a matter of luck but quite infrequent, however, attestations should be done once per epoch.
How we view rewards while considering risk in staking is a subject of research at Chorus One. This piece aims for other validators and interested parties to understand the “main-principles” they need to follow in order to minimize losses, and in turn, maximize profits in the process of validation. In our study, we found that currently the expected annualized reward for an ideal validator (perfect performance) is 5.44%. This amount decreases to 5.4% when we take into account a less idealized case.
After providing validators a feeling for how much they stand to earn, the following part will present a more practical example and explain how these figures may actually vary.
Slashing risk is a type of platform-dependent risk, as platforms that offer a similar service carry common risks. This section refers to the different types of penalties, and methods to calculate them at certain scenarios.
All formulas presented have been transformed in order to give a more general idea. The risk modeling was done using the actual definition from the Beacon Chain specs (Phase 0 and Altair) and the state of the chain at the time of writing. More on our methods can be found in our full study, linked previously.
Slashing includes all penalties that result in the partial or total loss of the staked assets of a validator, which range from 0% to 100% of the assets. Failing to perform the current validator duties properly (see last section) leads to being penalized and, in the case of slashable actions, being forcefully ejected from the Beacon Chain for suspected malicious activity. This is done to protect both the validator from further losses, and to help the chain finalize.
The main reasons from slashing that validators must be aware of are: proposing two different (conflicting) blocks, submitting two different (conflicting) attestations and submitting an attestation that completely surrounds or is surrounded by another attestation. If these events are not the result of a malicious action, then it follows that they must come from a bug or error. To account for this, the amount of staked destroyed is proportional to the number of validators slashed around the same time. If this number is small, then it is unlikely to be the result of a coordinated attack because that would require a high number of validators. These “honest mistakes” are punished lightly, at a minimum of 1ETH. If, on the other hand, there’s a high number of validators slashed at the same time, then it is assumed to be an attack, resulting in a higher amount of stake being destroyed, up to the full balance of the node.
There’s of course a certain pressure on validators to avoid going out at the same time as other validators. This expectation to decentralize touches on aspects of client diversity, but also on sources of truth or hosting for clients. This is a very critical point for everyone participating in the Ethereum ecosystem, and one that Chorus One has considered in our design. But back to the topic, these penalties hold true whether or not blocks are being finalized (meaning, 2/3 of validators weighted by stake are online and their votes are being counted). This is the state of normal operations for Ethereum. Anything under that, and we can no longer reach agreement and the inactivity leak mentioned previously comes in to restore balance.
With a clear understanding of the rewards system, estimating the source of possible penalties is much more simple, by calculating the attestation reward/penalty delta.
Indeed, if the reward is not finalized, the corresponding amount is removed from the attester’s balance (the minimum penalty). At that point, the validator’s unlock date for the stake is delayed about 36 days. This is to allow another, potentially much greater, slashing penalty to be applied later, once the chain knows how many validators have been slashed together around the same time (further penalty). If an inactivity leak is active, then the potential reward drops to 0, so by fulfilling the duties you are only able to avoid penalties.
Since getting the source vote wrong implies getting the target vote wrong, and getting the target vote wrong implies getting the head vote wrong, the possible slashing scenarios reduce to these:
To quantify the outcomes for performing validator duties, we would like to compare what could be considered a generic validator across a selection of edge scenarios. This example takes into consideration the following values:
Perfect Validator
To start off, let’s look at what this validator would earn if they, and all other validators, had an ideal participation record under the defined specs.
Attestations can be rewarded with a portion of the “base reward” for each of the correlated duties, weighted by the specific service provided. In the latest specs, the target vote receives the highest rewards, as it is the most important to reach consensus. The base reward is a constant across the network at all times.
BASE REWARD (in Gwei) = Effective balance * [Base reward factor / sqrt(staked ETH balance)]BASE REWARD = 32,000,000,000 * [64 / sqrt(10,000,000,000,000,000)] = 20,480 Gwei = 0.00002048 ETH
Following the upgrade, the block reward allocated is now ⅛ of total rewards as intended by the Ethereum researchers, rather than ⅛ of ¼ of rewards, as was the case pre-Altair. You may notice the delay reward is missing. Now, all attestations are given specific inclusion deadlines to claim their rewards in a gradual pattern, so prompt voting is logically accounted for.
Since all validators are supposed to attest at least one time during an epoch (for a perfectly working network), the number of attesting validators is equal to the total active validators divided by the number of slots per epoch.
ATTESTING VALIDATORS = ACTIVE VALIDATORS / SLOTS PER EPOCHATTESTING VALIDATORS = 300,000 / 32 = 9,375 validatorsTOTAL REWARD = BASE REWARD * ATTESTING VALIDATORSTOTAL REWARD = 20,480 * 9,375 = 192,000,000 GweiBLOCK REWARD = TOTAL REWARD / 8 = 24,000,000 Gwei = 0,024 ETHTARGET REWARD = 26 * TOTAL REWARD / 64 = 78,000,000 Gwei = 0,078 ETHSOURCE REWARD = 14 * TOTAL REWARD / 64 = 42,000,000 Gwei = 0,042 ETHHEAD REWARD = 14 * TOTAL REWARD / 64 = 42,000,000 Gwei = 0,042 ETH
Sync committees rotate rather slowly (every 256 epochs or every day), and validators selected can earn the sync committee reward for each slot in which they participate. A high number of validators will not be actually selected for this reward in a year’s time.
SYNC COMMITTEE REWARD = 2 * TOTAL REWARD / 64 = 6,000,000 Gwei = 0,06 ETH
Finally, we see the maximum possible reward in an epoch (this number also coincides with the minimum penalty for being offline or failing to fulfill the previous duties):
MAXIMUM REWARD = (BLOCK+TARGET+SOURCE+HEAD+SYNC COMMITTEE) REWARDMAXIMUM REWARD = 0,246 ETH
It is important to note that there’s still a potential variation to this reward of a few percent over the course of a year due to sheer luck (e.g. the probability of being chosen to propose, be in the sync committee, being offline exactly at the moment you are selected, etc. This applies even considering this ideal case, where the validator performs all their duties perfectly. The effect increases as the validator set grows, due to probability. Although not worrying in terms of an investment risk (marginal differences should even out over the course of a year), it still should be kept in mind as we delve into actual performance of validators in the network.
If we were to expand this timeline to a year, then the expected reward for this single validator per year sits at around 1.7428 ETH, which corresponds to the 5.44%% APY we mentioned in the previous section. A validator can optimally earn one base reward per epoch over a long time horizon.
Realistic Validator
However, to get a bit closer to model real-world rewards, we must consider the impact of a less-than-perfect validator performance.
As we learned previously, rewards are maximized for validators the better the network behaves. This helps disincentivize malicious behavior but also means that rewards can be reduced by external factors. A model that considers all the reasons why this validator might fail to produce attestations, produce blocks, or fail to propagate is an option, but here we wanted to observe: what would happen if we assume that 99.25% (a fair figure in reality) of active validators are actually attesting blocks? Also, we wanted to make a more conservative choice and assume that our validator was online 99.9% of the time.
As we can see, in this new realistic scenario, the total distribution shifts slightly. The expected annualized reward suffers a reduction of about 0.8% and the resulting expected APY reduces to 5.4%, as we had mentioned. The probability of certain events happening plays a huge part in this scenario, so this is just a starting point to analyze.
Slashed Validator
Next we wanted to estimate what would happen if our validator was caught committing a slashable offense, the ones that were previously outlined to result in substantial loss of stake. To do this, we will assume that we simultaneously sign two different blocks with 1000 validators. In this case, we suffer three types of slashing for each validator involved:
This corresponds to a total loss of -0.8282 ETH. It is worth noting that this slashed amount increases with the number of validators that are slashed at the same time, as was discussed in the slashing overview.
PoS Ethereum is a highly complex and elaborate system. It is thoughtfully designed, but can be difficult to fully grasp from a validator’s perspective, which can make staking seem like an uncertain and unpredictable endeavor. There is still a compelling amount of ETH that is yet to be staked, therefore we must all prioritize network participation and security in the future.
To make sense of the opportunities presented by staking, we wanted to explore the risks native to Ethereum and how said risks stack up versus the offered rewards after the Altair upgrade. Hopefully this article has helped to clarify why and how these rewards can vary from small changes in state or from bigger events, and also show how staking is a profitable business to take part in the long term. As we found in our analysis, the profit of a single validator is around 1.7428 ETH per year, or if we prefer to see in terms of percentage, this corresponds to a 5.44%% APY.
Based on the analysis performed, we find that the most realistic impact is low enough for self-cover to be a viable option, but not low enough to be completely trivial. We have identified the most relevant scenarios to come up with this conclusion. Additionally, we have found that risk can be significantly reduced by non-financial actions, such as promoting validator diversity and operator distribution, as well as putting in place mechanisms to maintain high validation quality standards. You can read our full report here.
Taking Ethereum from the individual to the masses will require a set of tools that accelerate the process of setting-up a validator whilst maintaining the same level of security and protection. At Chorus One we are working to make this a reality through our infrastructure services, and we are preparing to launch new services in the near future that take this to the next level. To learn more, please reach out to research@chorus.one.
.png)
