Umberto | Mostly Data 🧙‍♂️

Scheduler debates on Solana miss the point: no scheduler is best everywhere. Streaming wins for mostly independent flow. Short batching can win for time-coupled races. But those races are account-local, so applying the same rule to every tx is not welfare-optimal. The right question is not “batch or stream?” It is: where is contention, and what source of loss dominates there? That points to adaptive short-window scheduling — not a one-size-fits-all rule.

@KamBenbrik More you use AI more they can train and improve the model. It’s a do ut des…nothing at this world is free

The issue with the current AI economics is that you pay for a fixed subscription, but the more you use the AI and the more it costs the company. Curious about the distribution of usage / user on a specific tier. Users who don’t use AI are probably compensating for heavy users

@moonshiesty > so DAG gives better censorship resistance Do you mean MCP, right? Even if I think they are equivalent. For DAG is the ordering that in principle can be manipulated. Re the encryption, I guess that’s the issue. But I’m not an expert

i see, so DAG gives better censorship resistance what ive heard is that if you broadcast to enough concurrent proposers its unrealistic that a single proposer can meaningful change tx inclusion fairness but it sounds MCP has better guarantees there > A validator sees tx T1 on a shared object, but T1 is economically bad for them if the DAG encrypts txs is the difference is that DAGs still require object metadata, and so they leak more information to proposers so they can infer whether a tx is economically bad?

question: if blockchains with DAG based consensus (hedera, sui) haven't found success with trading, why would solana MCP be any different? only meaningful difference between MCP and DAGs is the batch auction ive never heard anyone active on DAG blockchains tell me they want a batch auction and never heard any MM or trader tell me they want to see solana look more like hedera or sui

A DAG can look similar to MCP when it is truly multi-proposer: multiple validators can add concurrent blocks, and final ordering is derived from that shared structure. Sui is closer to this model. Hedera is not: it is an event-DAG with virtual voting, not MCP-style batch inclusion. The key difference is tx inclusion fairness. Let’s comsider this Gedankenexperiment: A validator sees tx T1 on a shared object, but T1 is economically bad for them. They omit it from their DAG block and include T2 instead, which touches the same shared object and benefits them. Another validator includes T1 later. The DAG consensus then produces a valid order, but that order already reflects a strategic inclusion decision upstream. So DAG is different from MCP. MCP guarantees that any proposed in time tx is included, and ordering cannot be manipulated by proposers (in case of encryption, w/o you can still do things around that)

@MostlyData_ DAG based systems allow traders to get inclusion from multiple regions *not a consensus expert

@LoganJastremski Can you specify what you mean? The 150ms latency is on block finalization, which is a net positive outcome regardless of how you order txs (and until we are at >150ms slot time the two remains largely uncorrelated)

Fair correction! You’re right that under FixedBatch, the scheduler buffers everything regardless of conflict status. The parallelism is downstream of the scheduling decision, not a bypass around it. Non-conflicting txs still eat the batch delay even though they don’t need serialization. I overstated that. Stepping back from the intra-slot optimization for a second. What are your thoughts on 150ms latency (post Alpenglow) cost on non-contentious state (payments, etc) actually mattering since traditional finance settles on T+1 to T+2.

Lately I've spent a bit of time into data to measure real scheduling effects. However, some theoretical work was needed as well. Here you can see how: - Waiting up to the end of the slot is not good - Under competition, batching is the scheduler that minimize agents' loss - Despite what people think, streaming txs is the scheduler that reduce agents' loss in a higher variety of situations - Batching is bad when you are e.g. trying to land oracle updates

@joaomendoncaaaa Yeah, probably I should pivot again to make sense of how Universe works

Umberto, the mostly data legend, is trying to make sense of solana drama good luck Umberto

@0xSharples That’s interesting. Wondering why…probably one should deep dive into it

@teddyfuse You should definitely go to Naples and try zeppole there

Happy St. Joseph's Day! I hope many of you get to enjoy some delicious zeppole today.

@aceddeca1 @ChorusOne Most of the work is based on simulations, so not specifically for this that's more theoretical. But I have to think about empirical measures. Thanks for the support 🙏

@MostlyData_ @ChorusOne I meant more as in if you segmented blocks out by those using dz edge if that would give less noise in evaluating the metrics. Enjoying the article and trying to read in between work pauses. Appreciate all that you did to put it together. Has been educational

I don't think at this stage. DZ can in principle make you see a tx earlier, but I doubt this can change scheduler logic. For execution ordering, that's more interesting. The metric developed still is an on-chain fingerprint that remains a function of orderflow. DZ can change orderflow shape, so in principle this is an interesting case to study. Tho, at this stage I don't think you'll see meaningful differences

@ChorusOne Does DZ edge create an ideal set to evaluate Execution Ordering and Scheduler Logic?

@LoganJastremski (btw, thanks for reading it and providing feedbacks)

> On Solana’s parallel VM, non-conflicting transactions execute simultaneously regardless of scheduler design. They never queue behind each other, so the scheduling question is mostly moot for them That's assuming you don't delay txs execution. These execute in parallel only when scheduler decide the deadline of waiting ends. Otherwise these wait as any other tx. > These are exactly the systematic competition scenarios where your Figure 9 shows batching Pareto-dominates streaming. The point is subtler than this. If you account for the previous point (e.g. FixedBatch still make non conflicting txs waiting) you have that FixedBatch is only benefitting this type of txs. > So the SVM already produces something close to the adaptive batch sizing your conclusion calls for through the runtime itself. Non-conflicting txs get streaming via parallelism well, this is not true. Most of the confusion arise from your first assumption around non conflicting txs.

@barrett_io I looked at who rt'd it...who is supposed to be the "mev ppl whose job it is to take your money"?

this is is being rt’d by mev ppl whose job it is to take your money think for a second

Most people sleep on how much schedulers affect Solana's transaction execution — and even more misunderstand what they actually do. Spoiler: execution order alone doesn't create value. Full breakdown in the article 👇

@alessandrod I now want the other 2

top 3 best article nobody will ever read

💯 for sure if you measure that Solana is the leading ecosystem Was mostly complaining about interpretation from volume. To me, it’s still not a suitable metric. For example, you see uniswap as 2nd, well if you check for volume most of it is coming from sandwiching (where you need to move a lot of liquidity to steal few $). Is looking at volume informative for retail?

@MostlyData_ @therealchaseeb Volume is a proxy measure and a crude one at that but it does tell an important story - people, mostly retail, want to trade and use applications that are completely native to Solana. That's important.

Just waiting till the rest of world realizes that retail activity is literally all that matters and this entire industry has been focused on the wrong user demographic for years

@ramzyyalii @therealchaseeb Sure, but lot of snipes, manipulation etc. So to me volume is not a clear metric to evaluate it.

@ramzyyalii @therealchaseeb The assumption is that all volume on pump is retail?

@LoganJastremski @0xrafal You may enjoy this read x.com/mostlydata_/st…

Solana’s MCL isn’t implementing Budish’s FBA mechanism literally The anti adverse selection benefits come through a different pathway via leader competition, application-level maker protection with PropAMMs and cancel-first queues These aren’t identical to uniform-price batch clearing, and they may not capture 100% of the theoretical idea but they address the same root problem single entity monopoly over transaction ordering creating extractive rents through architectural competition

The high-frequency trading arms race is a symptom of flawed market design. Instead of the continuous limit order book market design that is currently predominant, we argue that financial exchanges should use frequent batch auctions: uniform price double auctions conducted, for example, every tenth of a second. That is, time should be treated as discrete instead of continuous, and orders should be processed in a batch auction instead of serially. Our argument has three parts. First, we use millisecond-level direct-feed data from exchanges to document a series of stylized facts about how the continuous market works at high-frequency time horizons: (i) correlations completely break down; which (ii) leads to obvious mechanical arbitrage opportunities; and (iii) competition has not affected the size or frequency of the arbitrage opportunities, it has only raised the bar for how fast one has to be to capture them. Second, we introduce a simple theory model which is motivated by and helps explain the empirical facts. The key insight is that obvious mechanical arbitrage opportunities, like those observed in the data, are built into the market design—continuous-time serial-processing implies that even symmetrically observed public information creates arbitrage rents. These rents harm liquidity provision and induce a never-ending socially wasteful arms race for speed. Last, we show that frequent batch auctions directly address the flaws of the continuous limit order book. Discrete time reduces the value of tiny speed advantages, and the auction transforms competition on speed into competition on price. Consequently, frequent batch auctions eliminate the mechanical arbitrage rents, enhance liquidity for investors, and stop the high-frequency trading arms race.