The join optimization cost model uses onDiskSizeBytes, computed as 

recordStore->storageSize() - recordStore->freeStorageSize()

 to estimate the number of on-disk pages for each collection. These values come from WiredTiger statistics (WT_STAT_DSRC_BLOCK_SIZE minus WT_STAT_DSRC_BLOCK_REUSE_BYTES). Even when restoring identical data via mongorestore --maintainInsertionOrder, WiredTiger's page layout is non-deterministic for a number of reasons:

• Page split points depend on cache pressure and eviction timing
• Checkpoint timing relative to insert batches is wall-clock driven
• Block allocation depends on the path-dependent free extent list.

This causes onDiskSizeBytes to vary by small amounts between otherwise identical runs even on the same machine.

We accounted for this variation as part of SERVER-122592 in the estimation of the number of pages in the wiredtiger cache by rounding onDiskSizeBytes / pageSizeBytes to the nearest power of 2^(1/4), giving a maximum of 9.5% error.

However, three call sites in join_cost_estimator_impl.cpp: costCollScanFragment(), costIndexScanFragment(), and costINLJ(), computed onDiskSizeBytes / pageSizeBytes directly without quantization. These unquantized values fed into the Mackert-Lohman random I/O estimator, causing the cost model to produce slightly different cost estimates across runs. When cost estimates for competing join strategies are close, this variation is enough to flip the winning plan.