Marginalia Search Engine Software Documentation

4.3 Crawl Data, Processed Data, etc.

The system stores crawl data in index-n/storage/, along with processed data and other various long-term data.

The data generated by the system is in general viewable within the control GUI, under Index Nodes -> Node N -> Storage.

Listing of data

Clicking on the paths in this view will bring up details.

Data details screenshot

This view will show the path of the data relative to the node storage root (e.g. install-dir/index-1 on disk), its relation to other data (e.g. crawl data to processed data), and any files in the root directory of the crawl data.

These can be directly downloaded from this view. Note that directories are not shown in this view. This is due to size constraints.

A special file marginalia_manifest.json exists in all storage objects. This enables the system to recover information about the directory in the event of an irrecoverable database error. It also means if you move crawl data from e.g. index-1 to index-2, then the system will pick up the change automatically.

In The Filesystem

In general, the names of the storage directories reflect their content. Crawl data is stored in directories with names like crawl_data_..., and so on.

Crawl Data

In the root of the directory is a file called crawler.log, which is a journal of what has been crawled, and where it is stored. It will relate domain names to subdirectories.

To get around file system limitations, the crawl data is stored in a subdirectory tree. Each domain is stored in a separate parquet file.

Crawling a sample set of marginalia websites will result in a crawl_data_…-directory with these files:


If we inspect the crawler.log file, which acts both as a journal for the crawler, and an index for navigating the data, we find these lines:

# Starting WorkLog @ 2024-02-12T16:38:47.267678766     2024-02-12T16:38:48.074759835   /storage/crawl-data__24-02-12T16_38_44.774/76/7a/   1      2024-02-12T16:38:49.258043503   /storage/crawl-data__24-02-12T16_38_44.774/50/07/    1    2024-02-12T16:38:49.258399009   /storage/crawl-data__24-02-12T16_38_44.774/37/0a/  1      2024-02-12T16:39:19.524394369   /storage/crawl-data__24-02-12T16_38_44.774/3d/44/    31       2024-02-12T16:42:12.546372348   /storage/crawl-data__24-02-12T16_38_44.774/69/9b/       194

Parquet files can be interrogated with SQL using e.g. duckdb.

Processed Data

Processed data follows an analogous pattern, with the difference that it does not construct a directory hierarchy, but stores its data directly in the root in much fewer files.


The processor.log file also has a different pattern, e.g.

#  Format: 
#  + ID signifies adding an item to the current batch
#  F    signifies finalizing the current batch and switching to the next
#  X    discard contents from the current batch and start over, written after a crash
# Upon a crash, items that have re-process until their batch is finalized
# 2024-02-12T16:43:07.789999695
# finalizing batchNumber = 0