Plan a shred-to-land latency budget

6 rows expose the whole path

The latency-budget planner gives detection, decode, decision, signing, submission, and land their own rows. Each row accepts low, typical, high, start boundary, end boundary, sample window, and evidence source.

The output shows local processing, end-to-end time, slack against an opportunity deadline, and the stage contributing most variance. It prevents a sub-millisecond decode claim from standing in for transaction execution.

Method

For planning ranges, let L_i, M_i, and H_i be the low, typical, and high estimate for stage i.

low_total = sum(L_i)
typical_total = sum(M_i)
high_total = sum(H_i)

typical_slack = opportunity_deadline - typical_total
high_slack = opportunity_deadline - high_total

Positive slack means the modeled path completes before the deadline. Negative slack shows the amount by which it misses. The component calculates local total through submission separately from total including land.

Adding p95 values does not produce a statistically valid end-to-end p95 unless dependence and distributions meet strong conditions. The simple sum is labeled a conservative planning view. When raw samples exist, join stage timestamps for the same event and calculate the end-to-end distribution directly.

If only separate empirical distributions exist, Monte Carlo can draw one value from each and sum them. Independent draws are a model assumption. Shared load can correlate decode, decision, and submission tails, so the tool accepts paired samples or a correlation warning.

Define every boundary

Detection might begin at kernel packet arrival and end when a candidate program match exists. Decode might end when all transaction fields required by the strategy are available. Decision ends with an immutable action or rejection. Signing begins after that action and ends with signed bytes.

Submission can end at local socket send, relay acknowledgement, leader acknowledgement, or another named point. Those are not interchangeable. Land ends when the transaction is observed at processed, confirmed, or finalized commitment under the user's definition.

The page rejects unnamed latency rows. A number without start and end boundaries cannot be reproduced.

Worked arbitrage example

Assume a desk enters these typical values:

detection = 0.6 ms
decode = 1.2 ms
decision = 0.8 ms
signing = 0.15 ms
submission = 5.0 ms
land = 180.0 ms
opportunity deadline = 250.0 ms

Local time through submission is:

0.6 + 1.2 + 0.8 + 0.15 + 5.0 = 7.75 ms

Total through land is:

7.75 + 180.0 = 187.75 ms

Typical slack is:

250.0 - 187.75 = 62.25 ms

The same plan uses high values of 1.0, 2.0, 1.5, 0.3, 12, and 450 milliseconds. High total is 466.8 milliseconds. High slack is negative 216.8 milliseconds. The strategy fits at the typical modeled path and misses badly in the high path.

That result points to landing variance, not decode. Cutting decode from 1.2 to 0.6 milliseconds adds 0.6 milliseconds of slack. Cutting typical land from 180 to 120 adds 60 milliseconds. The table makes optimization priority visible.

Opportunity deadline needs evidence

An opportunity lifetime can be measured from signal-ready time until expected value crosses zero, a price limit fails, or a competing state change lands. It is not always one slot. Different sizes and routes can have different deadlines.

Record the outcome after each submitted event. Group opportunity decay by strategy, market, leader, fee, time of day, and contention where sample size permits. A fixed 250-millisecond deadline can be a useful first model and a poor production model.

No public Solana mempool supplies a universal pending-to-inclusion timer. A shred observation is already in a proposed block. Reaction time begins at that observation, while the response normally competes for a later position.

Separate controllable and external stages

Detection, decode, decision, and often signing are mostly controlled by the user's host. Submission crosses a service or network boundary. Landing depends on leader scheduling, policy, account locks, priority fees, compute, validity, competition, and forks.

Coloring stages by owner prevents the wrong optimization. A team can tune receive queues while also recognizing that a distant submit path dominates. It can choose a slower signer for security and see the exact slack cost.

The Frankfurt feed origin affects arrival geography. It says nothing about the next leader or outbound submission path. Measure both directions rather than treating region as a constant latency deduction.

Import measured events

CSV import accepts event ID and timestamps for packet arrival, candidate detection, decode complete, decision complete, signing complete, send, acknowledgement, landed observation, and commitment. The component checks monotonic order and reports missing stages.

Event IDs matter because sums across unrelated medians can describe no real event. Joined samples show whether packet bursts create decision and submission tails together. Report p50, p95, p99, maximum, sample count, and capture dates.

Use monotonic clocks inside a process or host. Cross-host stages require synchronized clocks and an uncertainty estimate. Negative stage times usually reveal clock or join errors and remain visible rather than being clamped to zero.

Failure paths consume a budget too

Add deadlines for missing-shred recovery, remote signer timeout, submission retry, blockhash refresh, and reconciliation. A retry can improve land probability while extending age or duplicating exposure if identity rules are wrong.

Rejections should have latency histograms. A risk check that times out after 100 milliseconds can block safer candidates behind it. Fail closed within a bounded interval and keep unrelated markets independent.

Expected value can combine land probability and outcome, but the planner does not turn milliseconds into profit without user inputs. A faster path with higher failure cost can be worse.

Limitations

Planning ranges are not guarantees. Low and high values depend on hardware, geography, program mix, packet loss, recovery, signer policy, leader path, and network state. The example is illustrative.

Summing percentiles can overstate or understate the true end-to-end percentile. Joined production samples are preferred. Monte Carlo output inherits its distribution and correlation assumptions.

Seeing a shred first does not guarantee landing first. The tool quantifies that distinction but cannot predict leader admission, competing private routes, forks, or strategy profit.

The planner does not send transactions or benchmark the user's host. Its output is an explicit hypothesis. Instrument the actual boundaries and replace every estimate with a dated measurement.