E_DATA_003 — Aggregator cannot run on partitioned workers

BinaryLogReader::run(aggregators, n_threads) (DataReader.run in Python) splits each segment at compressed-block boundaries and gives every worker a clone of the panel. Some aggregators cannot be cloned that way: BookSnapshotBinAggregator rebuilds the order book by applying deltas in order from the start of the tape, so a worker that begins in the middle has no book to apply them to.

The reader asks each aggregator through IAggregator::supportsParallel() before it partitions anything. If the panel holds one that says no and the caller explicitly asked for more than one thread, the run stops here, before the first byte is read.

With n_threads=0 (the default, "auto") there is no error: the reader resolves the run to a single thread. Earlier versions let the request through and threw from inside a worker partway along, which made the outcome depend on whether the tape happened to be compressed into enough blocks to trigger partitioning at all.

How to fix

Ask for one thread when the panel needs the whole tape in order.

PythonC++

reader = lrvx.DataReader("./tape")
reader.run([book_agg], n_threads=1)

lrvx::replay::BinaryLogReader reader(cfg);
std::array<lrvx::replay::IAggregator*, 1> panel{&book_agg};
reader.run(panel, 1);

Or split the panel: run the order-dependent aggregator on its own single-threaded pass and let the rest go parallel. Two passes over a compressed tape usually still beat one single- threaded pass over everything.

Writing an aggregator that says no

Override supportsParallel() when the result depends on event order across the whole tape:

class MyBookAggregator : public lrvx::replay::IAggregator
{
 public:
  bool supportsParallel() const override { return false; }
  // cloneEmpty() / merge() may still throw as a backstop.
};

The default is true, so aggregators that accumulate independently per event (counters, histograms, OHLC bins) need no change.