Poker Platform, Scalability, Vendor Due Diligence, GLI-19, PokerEnginePro
By Cyntheion Research
2026-10-31 · White Paper · 7 minutes
Short answer. At scale, a poker platform must keep each table independent, add capacity on its own, upgrade without stopping play, account for every chip and replay every hand. Its random number generator must also be ready for laboratory certification. Ask any vendor for measured evidence of each one, with the test conditions stated, not just design targets.
- PokerEnginePro has been measured at 50,000 concurrent players across 10,000 tables, on one development machine.
- In that run the slowest 1% of updates (p99) reached players within 85 ms, against a 250 ms budget, with zero chip errors.
- Losing a server mid-hand voids only that one hand and refunds it exactly once. Upgrading a server voids none.
- 11.3 million ledger entries across about 990,000 accounts were reconciled in 28 seconds, with no drift.
- The design is sized for one million registered players. That is a design target, not a measurement.
Last updated: 31 October 2026 · Cyntheion Research
In this article
- Why is poker harder to scale than most apps?
- What must a platform guarantee at scale?
- What has been measured, and what is designed for?
- What should you ask a platform vendor?
- Frequently asked questions
- How PokerEnginePro fits
Why is poker harder to scale than most apps?
Every poker table holds live state that moves real money every few seconds and must happen in strict order. A slow page on a shop is an annoyance. A raise processed before the fold it followed is a wrong hand, and a regulator will ask why.
Four properties set the bar:
- Order. Actions at a table form a single sequence that can never be reordered.
- Money. Every chip that enters a pot must leave it to a winner or the rake, and a retry must never pay twice.
- Fairness. The deal must be unpredictable, unbiased and testable by an independent laboratory.
- Uptime. Players sit for hours. Peak traffic and release days both arrive while hands are in play.
What must a platform guarantee at scale?
A platform you can grow on makes six guarantees and can prove each one with a test result. Here they are, with what PokerEnginePro has shown for each.
1. One fault never stops the room
Each table runs on its own, so a problem at one table or on one server cannot freeze the others. When a server is lost mid-hand, its tables carry on elsewhere. The interrupted hand is voided, every stack is restored, and the refund is paid exactly once.
In failover tests, killing a server mid-hand voided one hand. Draining a server for maintenance voided none.
2. Capacity grows with traffic, automatically
The platform should add capacity when players arrive and release it when they leave, without staff on call. In a production-like test environment, the connection tier grew from 2 to 4 servers as load rose and returned to 2 when it fell. That environment passed 111 of 111 checks with zero restarts.
3. Upgrades happen during play
A release should never need a maintenance window. A server being upgraded finishes the hands it is running, starts no new ones, then hands over. Old and new versions run side by side during the rollout, and players who reconnect pick up where they were.
Losing a server should cost one hand, refunded once. Upgrading one should cost nothing.
4. Every chip is accounted for and reconciled
A double-entry ledger records every movement of money twice, as a debit and a matching credit, so the books always sum to zero. Every hand should balance on its own, and every money command should be idempotent: if a request is sent twice, it takes effect only once.
PokerEnginePro checks chip conservation after every state change. Its nightly reconciliation covered 11.3 million entries across about 990,000 accounts in 28 seconds and found no drift. The ledger and wallet page explains balances, escrow and rake.
5. Every hand can be replayed for disputes
The same cards and the same actions should always produce the same hand. That makes every hand replayable for a player complaint, a regulator or an integrity review. Players should never see a card they were not entitled to see, even in a replay.
PokerEnginePro replays 1,500 random hands at every decision point on each test run and requires 14 reference hands to match byte for byte. See the game engine overview.
6. The RNG is ready for certification
Regulated markets require an independent laboratory to test the random number generator. GLI-19, the Interactive Gaming Systems standard from Gaming Laboratories International, is now at version 3.0[1]. A platform should give the laboratory a traceability matrix, game rules, an RNG sample harness and a test plan.
PokerEnginePro is built for GLI-19 and its certification kit is ready. Laboratory certification is on the roadmap. Every deck is committed before the first card is dealt, so a player can check the cards they saw afterwards in their own browser. Our RNG certification guide covers what laboratories ask for.
And it should run where you need it. Some markets require servers in the country. PokerEnginePro is designed to run on the major clouds or on your own hardware; packaging for some environments is still being completed. The cloud page shows where it stands.
What has been measured, and what is designed for?
A vendor's numbers are only useful when you know which ones were measured and under what conditions. PokerEnginePro's figures:
| Metric | Designed for | Measured | Conditions |
|---|---|---|---|
| Registered players | 1,000,000 | Not measured | Design target |
| Concurrent players | 50,000 | 50,000 | One 16-core development machine, load generators on the same box |
| Tables in play | 10,000 | 10,000 | Same run |
| Hands per second | 200 | 232 | Same run |
| p99 update to players | Under 250 ms | 85 ms (131 ms on a repeat run) | Same run |
| Chip errors | 0 | 0 | Every hand, every run |
| Certified on reference hardware | 50,000 concurrent | Planned | With a server deliberately killed mid-run |
Put plainly, PokerEnginePro has measured 50,000 concurrent players, not a million. The million is the number of registered players the design is sized for. The scale page lists every measured figure.
What should you ask a platform vendor?
Ask for evidence, not adjectives. These questions separate tested platforms from brochures:
- Concurrency. What is the largest number of concurrent players you have measured, on what hardware, and for how long?
- Latency. What were the median and p99 times from a player's action to every other player's screen?
- Failure. What happens to a hand in play when a server dies? How many hands were voided in your last failover test, and how were they refunded?
- Upgrades. Can you release a new version while tables are running? How many hands did your last rolling upgrade void?
- Money. Is the ledger double-entry? Are money commands idempotent? How long does a full reconciliation take, and when did it last find drift?
- Disputes. Can you replay any hand from its stored record and get exactly the same result?
- RNG. Which standard is the RNG built for? Is it certified, or is it ready for certification? Can you show the laboratory kit?
- Hosting. Can it run in a specific country, on your cloud or your own hardware, with the same release?
- Targets versus results. Which of your published numbers are measured, and which are design targets?
Frequently asked questions
Has PokerEnginePro been tested with a million players?
No. One million is the registered-player design target. PokerEnginePro has measured 50,000 concurrent players on one development machine. A certified run on reference hardware is planned.
What happens to a poker hand if a server fails?
On PokerEnginePro the interrupted hand is voided, every stack is restored and the refund is paid exactly once. The tables carry on on another server. A server taken down for an upgrade finishes its hands first, so none are voided.
Is PokerEnginePro's RNG certified?
Not yet. It is built for GLI-19 and the certification kit is ready, including a traceability matrix and an RNG sample harness. Laboratory certification is on the roadmap.
What does p99 latency mean for a poker table?
p99 is the time within which 99% of updates arrive. In PokerEnginePro's 50,000-player run, 99% of updates reached every other player at the table within 85 ms, against a 250 ms budget.
Why does a poker ledger need to be idempotent?
Networks retry. An idempotent command takes effect once however many times it is sent, so a repeated payout or buy-in never moves money twice.
How PokerEnginePro fits
PokerEnginePro is a licensable poker platform for operators who need these guarantees with evidence behind them: independent tables, automatic capacity, upgrades during play, an exact ledger, replayable hands and a GLI-19 certification kit. Start with the scale page, review the security page, or talk to us about licensing.
Disclosure: Cyntheion builds and licenses PokerEnginePro. This article is general information, not legal, tax or financial advice.
Sources
- "GLI Standards" (GLI-19 Interactive Gaming Systems v3.0), Gaming Laboratories International, https://gaminglabs.com/gli-standards/, accessed 31 October 2026.
