| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Instead of having the asynchronous preprocessor concept with different
pipelines for new/modify/delete we have a single pipeline with
synchronous preprocessors that act upon new/modify/delete.
This keeps the code simpler due to lack of asynchronity and keeps the
new/modify/delete operations together (which at least for the indexing
makes a lot of sense).
Not supporting asynchronity is ok because the tasks done in
preprocessing are not cpu intensive (if they were we had a problem
since they are directly involved in the round-trip time), and the main
cost comes from i/o, meaning we don't gain much by doing multithreading.
Costly tasks (such as full-text indexing) should rather be implemented
as post-processing, since that doesn't increase the round-trip time directly,
and eventually consistent is typically good enough for that.
|
