The Second Pass Carried Every Frame

The file is 4,611 bytes and its sha256 begins c23c493c7cad6f6b. It holds 11 records, one for each still in the shoot. Every one of the 11 reads done. Inside those 11 rows sit 44 string values, and 24 of them are distinct, because the status is the same word 11 times and the model name is the same string 11 times. Two thirds of the prose in this file is one word repeated. The file is also smaller than a single one of the images it describes, by a factor of 74. The lightest still in it is 342,929 bytes. That is 74 times the size of the record that proves it exists, and that ratio is the entire job of a log.

Nothing in the file declares a schema, a column list or a version. The shape of a row is the schema. All 11 rows carry the same 6 fields in the same order: a status, the url the image came back on, the path it was written to, a byte count, a tick count and the model that made it. 11 rows times 6 fields is 66 values, and 44 of those are strings. Not one row deviates from that shape. That is the good news and the warning in the same breath, because a file whose rows cannot differ cannot tell you that two runs were different. It can only tell you that 11 things happened, which is the weakest thing a record can say.

Every record carries a tick count of 400,000,000. Exactly that, 11 times out of 11, which sums to 4,400,000,000 ticks across the file. The bytes, in the same rows, run from 342,929 to 560,710, a spread of 1.64. The mean is 383,271 and the median is 368,835, and the median is a real row in this file rather than an artefact of the arithmetic. So the payload moves 64 per cent between the lightest frame and the heaviest while the cost column reads identical in every row. A column that never moves is not a measurement. It is a constant with a field name, and it will hide the one render that costs ten times the others.

Every key is 38 to 44 characters long and every one opens with the same 22. After that prefix comes the shot name, then a dot, then the shot ordinal as two digits, then the version: V1.0. All 11 carry V1.0, and the ordinals run 01 to 11 with no gap in the set. The version also sits inside the folder path each image was written to, so it is written twice per record and declared as a field zero times. The folder holds 18 logs and they run V1.0, V2.0, V3.0, V3.1, V3.2 and V3.3. The batch beside this one is V2.0, has 20 rows, all done, 9,876,238 bytes declared, and carries 4 fields instead of 6. The schema shrank while the batch grew, and the only place the version lives is a string a human reads.

4,611 bytes. 11 records. 11 of 11 reading done.

One folder, two logs, two outcomes. The sibling file is 5,275 bytes, sha256 e9464138c10d8a33, and holds 22 records: 10 done, 11 failed, and 1 with no status at all. Its 11 failed rows carry a single 25-character error between them, written 11 times, and 18,194,412 bytes are declared by the 10 rows that worked. This file declares 4,215,981 bytes across 11 rows and has no error field at all. Both were written by the same pipeline into the same directory, days apart, and nothing in either one records which pass came second. The version token is the closest thing to an attempt counter in the folder, and it lives inside a filename, where no query can sort it.

The probability that a folder of logs like this can tell you, six weeks later, which pass was the retry is low, and most teams are pricing it at zero. The cost of fixing it is 1 field and the cost of not fixing it is a rewrite you cannot justify. Eleven frames, 4,215,981 bytes and a flat meter is a clean run with no story in it, which is exactly when a log should be doing its cheapest work. On 210 energy systems deployed across Africa and Asia, a unit was signed off only when the record, the serial and the site all resolved on the same day, and the failures went into the same ledger, in the same columns, with the same date. Write the attempt, the version and the cost where a machine can sort them, or the file you keep is a receipt rather than a record. Keep more of yourself, for longer.