Before increasing proxy capacity, put one atomic byte ledger in front of new work: reserve an estimated upper bound, admit only if the remaining local budget stays non-negative, then replace the reservation with authoritative actual usage. This prevents many concurrent workers from each treating the same remaining allowance as theirs. It does not make a supplier meter instantaneous, guarantee a bill ceiling, or turn a local response-body estimate into an IPHTML limit.
This guide is for an engineering lead or support owner running an authorized, continuing collection or QA workload who needs a conservative stop rule while usage from in-flight requests is still unknown. The ledger below is deliberately synthetic and hand-recomputable. It is an operational control for local request/response body bytes; it is not a supplier bill meter, a monetary hard cap, or a claim about IPHTML plans, prices, capacity, or performance.
1. Choose the accounting boundary before admitting work
Write down one accounting period, one unit, and one owner of the decision. This example uses MiB (1 MiB = 1,048,576 bytes) in a local period. Record the timezone and reset/close boundary too: an unresolved prior-period reservation must not disappear at midnight. Select the byte fields your client can reliably capture—for example, request and response body bytes after the application has received them—and state what is excluded. Headers, connection overhead, failed transfers, retries, compression, provider rounding, and traffic counted by a supplier may differ. Do not silently compare this ledger to an invoice until the period, units, and meter definition have been reconciled.
| Field | Meaning in this guide | Must not be interpreted as |
|---|---|---|
B cap | A local planning limit for this period, in MiB. | A provider quota, invoice total, or guaranteed spend limit. |
H buffer | A deliberately unavailable cushion for unmodeled or delayed information. | Guaranteed coverage for every meter difference. |
C settled | Authoritative locally observed body bytes for completed work. | Proof a supplier has billed the same amount. |
R outstanding | Reservations for admitted work whose actual bytes are not yet authoritative. | Traffic already proved to have occurred. |
At every decision point calculate available = B − H − C − R. Admit a request only when its proposed reservation q is no larger than available, and record the reservation with an idempotency key in the same durable transaction. If no finite reservation bound can be estimated from a measurable workload, HOLD until it can; do not treat the buffer as a substitute for a bound. A client-side check followed later by a separate write is not admission control: two workers can both see 350 MiB and each reserve 250 MiB.
2. Recompute a synthetic reservation ledger
Start with B=1000, H=100, C=400, and R=150 MiB. Available capacity is 1000 − 100 − 400 − 150 = 350 MiB. Worker A asks to reserve 250 MiB. One coordinator atomically writes A’s reservation, changing R to 400 and available to 100 MiB. A concurrent worker B asking for 250 MiB is therefore HOLD, not admitted.
| Event | Cap B | Buffer H | Settled C | Outstanding R | Available | Decision / evidence |
|---|---|---|---|---|---|---|
| Opening state | 1000 | 100 | 400 | 150 | 350 | Existing settled record and an earlier open reservation |
| A reserves 250 | 1000 | 100 | 400 | 400 | 100 | ADMIT; durable reservation ID A |
| B requests 250 | 1000 | 100 | 400 | 400 | 100 | HOLD; no reservation is written |
| A settles at actual 180 | 1000 | 100 | 580 | 150 | 170 | Authoritative local byte evidence replaces A’s 250 reservation |
| Unknown U remains reserved at 150 | 1000 | 100 | 580 | 150 | 170 | HOLD; absence of a completion record is not a refund |
| U settles at actual 120 | 1000 | 100 | 700 | 0 | 200 | Later authoritative local evidence |
| New job C reserves 180 | 1000 | 100 | 700 | 180 | 20 | ADMIT only through the same coordinator; this is not a smaller replacement for B |
Download the bilingual reservation-ledger CSV. It reproduces those rows as modeled values and deliberately leaves every operator actual and evidence reference cell blank. Its values are not observed traffic. Save the download as proxy-traffic-budget-control.csv, then use this standard-library-only check; it recomputes each numeric row rather than trusting the displayed available column.
import csv
with open("proxy-traffic-budget-control.csv", newline="") as f:
for row in csv.DictReader(f):
if row["cap_mib"]:
available = int(row["cap_mib"]) - int(row["buffer_mib"]) - int(row["settled_mib"]) - int(row["outstanding_mib"])
assert available == int(row["available_mib"]), row["event_en"]
print("ledger arithmetic checks")3. Settle once; do not double-count
A reservation is a temporary worst-case allocation, not an extra charge. When A’s authoritative local evidence says 180 MiB, move exactly that actual amount into C and remove A’s 250 MiB from R. Adding 180 without removing 250 double-counts; removing 250 and adding nothing undercounts. Store a settlement ID tied to the reservation ID, and make retries of the settlement operation idempotent so a delayed message cannot post the same 180 MiB twice.
Release a reservation only with evidence that the request was never dispatched or was canceled before dispatch. If it was dispatched but completion evidence is missing, retain the reservation as UNKNOWN; do not auto-refund it merely because a timeout, worker lease, or reporting deadline expired. A later local completion record may settle it. If no evidence arrives by the close of the period, carry the unresolved reservation into the conservative close/reconciliation process and document who owns follow-up.
4. Treat unknowns, deadlines, and overages as decisions—not zeros
| State | Ledger treatment | Operational action |
|---|---|---|
| Not dispatched, evidenced | Release its reservation once. | Record the cancellation or queue evidence and idempotency key. |
| Dispatched, actual missing | Keep the reservation in R. | HOLD new discretionary admissions; investigate the missing completion evidence. |
| Actual exceeds its reservation | Settle the real actual; available can become negative. | STOP further admissions, record the overshoot, and reconcile. A local stop cannot undo billing already incurred. |
| Period deadline passes | Do not transform UNKNOWN to zero. | Carry it conservatively or use an explicit close policy with an evidence trail. |
A negative available figure is evidence that the local control was exceeded, not permission to keep sending work because a dashboard has not changed yet. Cloud cost-budget documentation is a useful analogy here: alerts and cost data can be delayed, so a notification is not a real-time stop mechanism. AWS documents that limitation for its own cost tools; it does not describe IPHTML, proxy billing, or this worksheet. Likewise, Oxylabs documents supplier-specific limitation modes and units such as GB versus WebAPI requests. That confirms the need to read the applicable supplier terms; it does not establish any IPHTML setting or meter.
5. Decide whether to expand capacity only after reconciliation
Do not buy or enable more capacity just because an individual job was held. First ask whether holds came from a deliberately small local cap, a buffer chosen without evidence, an oversized per-request reservation, duplicate work, or unknown completions. The duplicate-request guide helps reduce unnecessary demand; the retry-budget guide sets per-job stop rules; and the workload-sizing guide separates transfer allowance from rate and concurrency. For a broader outcome view, use the reliability scorecard; none of those pages is replaced by this ledger.
- Keep the current cap: evidence is complete, holds are expected, and the authorized workload fits the agreed local policy.
- Improve measurement first: body-byte fields, dispatch evidence, or completion records are missing. More capacity would hide the accounting gap.
- Re-estimate reservations: completed comparable work consistently shows a documented difference, while retaining a conservative buffer and review date.
- Discuss a service requirement: prepare the accounting period, workload shape, local ledger, unknown-state policy, and applicable supplier terms. Do not send credentials or sensitive targets.
Bring an evidence-backed budget question to support
For an authorized workload, first assemble the local period, unit definition, request body/response body capture rule, peak concurrent reservations, unknown-state count, and the supplier terms you need clarified. Review IPHTML’s rotating proxy page to frame configuration questions, then use the official contact page for a specific service discussion. Confirm current terms and measurement definitions before making a buying or scaling decision; this worksheet promises no allowance, alert, price, performance, or billing outcome.
Method note: AI-assisted editorial analysis. Sources checked October 29, 2026: AWS Budgets: Managing your costs, used only for its delayed-alert/possible-overshoot analogy; and Oxylabs limitations documentation, used only as a supplier-specific illustration that units and controls vary. The numbers and ledger are synthetic. The local CSV arithmetic is checked offline; no live proxy test, supplier-meter reconciliation, customer outcome, payment, or renewal result is claimed.