A price chart makes a market look like one continuous number. An order book is less smooth. At any moment it contains a collection of offers to buy and sell particular quantities at particular prices. The last transaction tells us where two orders met. It doesn't promise that the next large order can trade there too.
That distinction is central to the ETH-USD event on GDAX on June 21. In the exchange's account published the following day, a large market sell order moved through the available bids. The initial decline triggered stop-loss orders and margin liquidations, adding further selling. The exchange said its matching engine operated as intended. That is its reported explanation, not proof that every surrounding system or product decision was satisfactory.
A small fictional order book makes the mechanism easier to see. Imagine buyers offer to take 10 units at 100, another 10 at 99, and 10 more at 98. A market order to sell 25 units consumes all of the first two levels and half of the third. The seller doesn't receive 100 for all 25 units. The order trades against the liquidity that actually exists.
Now add a rule owned by another participant: sell if the price falls below 99. A triggered stop that becomes a market order adds supply precisely when nearby bids are being consumed. A leveraged position may generate a similar effect when its collateral no longer satisfies the platform's requirements. The new selling lowers the price further, potentially activating more rules.
No individual rule needs to contain an infinite loop for the combined system to create feedback. Each participant reacts to a shared signal. Their reactions change that signal. The speed of the response matters because there may be little time for new buyers to place orders between successive events.
I think this is a useful reason to separate three questions. Did the matching engine follow its rules? Did customers understand what their order types could do? Was the surrounding risk system designed for the available liquidity? A yes to the first question does not settle the other two.
The chart also hides the difference between a brief extreme trade and a price at which a large quantity could actually change hands. A low printed on a screen is an observed event. It is not necessarily a broadly available opportunity. Reconstructing what happened requires quantities, order types and timing, not only the endpoints of a vertical line.
For me, the engineering lesson is about composing automated systems. A sensible local rule can become unstable when thousands of copies respond to the same input. Testing one rule in isolation won't reveal that behavior. The important question is what happens when its output becomes the next rule's trigger.