The solution I use is for the child processes to feed array updates back to the master process, which applies them to the array. This has worked for me because the worker processes I have needed thus far have always had fully-defined tasks at the moment they are forked — the only communication needed is to report their results back to the master process.

I have used this technique to maintain Tk GUI responsiveness while issuing network queries in the background, but the data to scrape is known when the process forks and the child only needs to parse a reply and pass the "important bits" back up.

That is very strange and I have not had that problem.

This sounds like you are not properly handling the stored objects. I use a single status pipe, where a child reports some identifier and an indication that a binary object is available, and a "return" pipe for each child, where the Storable data is actually written. The status pipe is non-blocking and monitored with Tk's fileevent mechanism, while the "return" pipes are used in blocking mode. I switched to this approach after trying to pass the Storable data back on the status pipe and finding that non-blocking mode caused problems, not to mention that the status pipe is shared, so the messages returned on it must be kept short enough to fit in the OS buffers to avoid messages from different children being interleaved.

Hey jcb!

Bit more details about my structure, and where the problem is:

- Perl/Tk with Tk::HList showing SQLite database view
- IPC::Shareable object "transport" for communication with "worker" child
- fork "worker" (forked before Tk objects become created) communicating with "transport" to Tk.
- "worker" creates its own IPC::Shareable "ac_trans" for its agents
-> On Tk's request (Button) the "worker" forks its agents which start doing the actual work.
- Agents download small bits of data (usually far away from 2^16 characters) and transfer it back to "worker" via "ac_trans".
- "worker" empties array within "ac_trans" and saves data to a regular array
- "worker" does this until all agents are exited
- "worker" fills database with all received data at once (faster than every agent accessing database for one small bit of data)
- "worker" informs Tk through "transfer" that work is finished.


Though: one source can't be predicted how many data will be downloaded and here it exceeds the 2^16 characters, IPC::Shareable can't handle, as well as Storeable can't (the way I tested it.).
So far I already have a database in use, I just create a table and use this for this only case. I think the more wait is worth it on this part. (SQLite supports cells with more than 2 billion characters.) Furthermore the changes are acceptable effort here.

perhaps have a look at Sys::Mmap and File::Map before using a disk-based DB just because it has locking mechanism.