2) Organize the log data into tables with columns, and then compress by column (so, each column has its own dictionary). This lets the compression algorithm perform optimally, since now all the similar data is right next to itself. (This reminds me of the Burrows-Wheeler transform, except much more straightforward, thanks to how similar log lines are.)

3) Search is performed without decompression. Somehow they use the dictionary to index into the table- very clever. I would have just compressed the search term using the same dictionary and do a binary search for that, but I think that would only work for exact matches.