What 24/7 Uptime Actually Means for a Multi-Asset Perps Venue

A trader sets a BTC-PERP stop at $78,000 on a Sunday afternoon and goes to a wedding. At 4am local time, two things happen in the next ninety seconds: BTC-USD on the big spot venues trades through $78,000, and the trader's perps venue goes dark on its primary price feed because an upstream exchange decided to roll a WebSocket connection on a weekend. Thirty seconds later the primary feed is back. What happened to the stop in the meantime?

That question is what "24/7 uptime" actually means. It is not whether the website loads. It is whether the matching engine still has a trustworthy mark price every single second of the week, and whether it refuses to match against a stale book when it doesn't.

Most venues that advertise 24/7 trading don't publish how they handle this. This post is a plain-English rundown of what the layers look like on JTX Markets, written so you know what to interrogate on any venue — ours or otherwise.

The three failure modes

Everything collapses into three categories.

• Feed source blinks. The upstream (Binance, Hyperliquid, Massive, oneZero, pick your flavour) has a connection reset, a brief rate limit, or an API outage. Could be ten seconds, could be ten minutes.
• The book goes stale. The engine still has a quoted bid/ask, but the quotes haven't been refreshed in long enough that trading against them would be matching against a time-traveller's book. On low-liquidity symbols this is the common one.
• The mark price drifts. Even if the book has recent quotes, the external reference price the engine uses for liquidations and funding has stopped updating. A trader with a position gets no mark-to-market, and risk calculations silently freeze.

A venue that only handles the first class of failure will quietly stop updating in the second two, and traders only find out when a liquidation fires against stale data or a stop sits unexecuted through a visible move.

What the layers look like on JTX

Four defences, each covering a different failure mode:

• Multi-source feed arbiter. Every tradable symbol has more than one price source. Crypto majors run Binance public WebSocket as the primary with Hyperliquid as the secondary; forex pairs come off a Massive REST poll with Hyperliquid as backup; equity perps mirror from a listed venue's futures contract. The arbiter runs a median-of-N filter with a sanity band — if one source disagrees by more than a configurable threshold, that source is excluded for the next tick and the median is recomputed from the others. One venue going dark is not the venue going dark.
• Per-symbol book-seeder watchdog. A separate process mirrors depth into the engine for every tradable symbol. If any one symbol's book goes stale for longer than 45 seconds, a watchdog fires and the symbol is auto-halted at the engine level until a fresh snapshot lands. Trading on a stale book is impossible, not just unlikely.
• Mark-price freshness gate. The engine rejects any order whose matching would reference a mark price older than 3 seconds. The number is tuned to be invisible under normal operation and strict enough that no fill ever happens against a frozen reference. If the mark truly stops updating, orders error cleanly rather than filling at nonsense prices.
• System-level health panel. The ops dashboard (visible to our team, mirrored into alerting) shows the per-symbol freshness state for every one of the 293 instruments. Any symbol with a mark older than 10 seconds or a book older than 30 lights up amber; anything stale beyond the halt threshold goes red and traders see a "paused" badge next to the symbol on the trade page.

These are the four layers between a feed hiccup and your account getting hurt. There is no single-point-of-failure design where one upstream disconnect takes trading offline for all 293 symbols; there is no quiet mode where the engine just keeps serving the last mark price it saw an hour ago.

What this means for the trader

A few specific implications you can test yourself.

• If a symbol goes "paused" mid-trade, that's working as designed. On any venue, a paused symbol beats a traded-against-stale-data symbol every time. Check how often this happens — on a mature venue you should see it rarely on majors, occasionally on long-tail symbols with thin upstream liquidity. If a venue never pauses anything, be suspicious.
• The spread widens when sources disagree. The arbiter's job is to keep the median honest, which means on a venue-divergence event the band around mark expands slightly. Your stops and TPs will still fire — reduce-only market orders don't care about spread — but a limit order that was sitting at the top of the book might sit through a widening without filling. This is correct behaviour; the alternative is filling into a bad quote.
• Liquidations use mark, not last-trade. When the mark price updates every 100ms from a median of multiple sources, your liquidation price is defended against a single wick on a single venue. Our matching engine uses mark for every maintenance-margin check; last-trade price only matters for execution, not for solvency.

Red flags to watch on any venue

If you are evaluating a perps venue beyond JTX, these are the questions worth asking:

• "What does your feed arbiter do when upstream sources diverge?" If the answer is "we use one source" or they don't know, you have your answer.
• "How long does a mark price have to be stale before an order errors?" Specific number is the right answer. "We monitor it" is not.
• "What auto-halts a symbol?" If nothing auto-halts, something going wrong requires a human to notice. Humans sleep.
• "Can I see the per-symbol freshness state as a trader?" On JTX this is on the ops system page for our team and reflected in symbol badges on the trade page. If a venue hides this, assume there's a reason.

24/7 is not a marketing claim. It is a specific set of promises about what the engine will refuse to do when a feed or a book or a mark gets ugly. The right answer to "we're always open" is a long one, not a short one, and if a venue can't give you the long answer, they probably don't have the layers that back the short one up.

If you want to see the stack running under live conditions, every symbol page carries a live freshness state. If you want to trade on it, open an account or start on a challenge tier and see how the system behaves across a weekend before you decide how much to size.