Planner Notes

From The Open Source Backup Wiki (Amanda, MySQL Backup, BackupPC)
Jump to navigationJump to search

This is a historical document, previously docs/multitape.txt. However, it's a great explanation of the planner and taper algorithms. Please take it all with a grain of salt, and consult the source as the final arbiter!

Introduction

The goal of this enhancement is to make Amanda independent of the number of tapes used per run or even per dump cycle. Specifically, I would like Amanda to handle the following:

  • output of amdump run goes to more than one tape
  • a single dump file can straddle two tapes
  • more than one amdump run can be done in a single day
  • planner should not care how many runs per cycle occur

And later:

  • multiple runs of amdump can go onto one tape (eg an append mode)
  • any dump files from a previous run that are on the holding disk are written to tape in this run (eg eliminate amflush)
  • taper write to multiple tape drives simultaneously


New Planner Algorithm

Time

Previously, planner marked time by the number of amdump runs, which it equated with number of tapes, and number of days. In Amanda 2.2, Amanda keeps track of the real passage of time, and doesn't generally care about the number of runs or tapes between any two events.

While Amanda 2.2 doesn't care about spacing between runs, dump cycles are still in terms of days, to make things easy to understand for the user. So, time differences are rounded to the nearest 24 hours:

  days_diff(A,B) = (<B> - <A> + 86400/2) / 86400

Where the times A and B are in seconds since the Unix epoch, and 86400 is the number of seconds per day. This rounds a 2.49 day difference down to 2 days, and a 2.5 day difference up to 3 days. No, Olafur, Unix time does not handle leap seconds. Give me a break. :-)

Full Backups

The first thing planner does is calculate when each filesystem is due for a full backup. This is trivial for normal backups:

full_is_due = days_diff(

There is a complication for "skip-full" filesystems. Under 2.2, these will be skipped on any runs that occur on the day the full is due, but we have to do the right thing if multiple runs are done that day, and if no runs are done that day (in which case we should be doing an incremental). Also, the time of last full dump is a fiction maintained by the planner -- Amanda has no way to tell whether the full backup was actually done or when it was done:

 	if(skip-full) {
 		if(full_is_due)
 			

Schedule Balancing

The runtapes parameter tells planner how many tapes it should plan to use each run. It multiplies this by the tape length to get the size available for the run. (NOTE: later amend this size if appending to tapes, or if there are dumps on the holding disk waiting to be flushed). Other than the size calculation, planner doesn't really care how many tapes will be written to.

The fundamental problem with attempting to balance the schedule is that we no longer know how many amdump runs will be done in a full cycle. The number may change from cycle to cycle if there are extenuating circumstances.

So, planner must guess at how many runs will be done in one cycle, by looking at the information for the last cycle, or, if this is the first cycle, assuming one run for each day in the dump cycle.

Overwrite Detection

When can a tape be overwritten, considering that it might have old dumps on it? We want to be able to warn when full dumps are going to be overwritten, but given the possibility of old files on the tape, how can we know when the tape is no longer needed? I think we can get this when going through the info file, considering each full dump and what tape it is on. Make sure we correctly handle stale information.

Taper Algorithm

Choosing a tape

taper must now handle writing to multiple tapes in one night, but choosing the tapes from the tape rack is done one at a time as needed, re-applying the same algorithm each time.

End of tape handling

As in earlier versions of Amanda, taper itself does not try to restrict writing to the tape size given in the config file. It relied on planner having correctly estimated backup sizes and limiting itself to what would fit on one tape.

Now, taper needs to switch to a new tape when the current tape has filled up. The tape is considered full when taper gets a write error. This will most likely occur in the middle of writing a (potentially large) backup file, perhaps even from a direct-to-tape socket, so there is no possibility of starting the backup file over again on the next tape, it must start from where it left off, rewriting the block that got the error on the next tape.

To insure correct operation, the file header of the continued file should contain an indication that it is a continuation, and at what offset. amrestore of course needs to be aware of this scheme and handle it correctly, perhaps by double-buffering internally. XXX provide more alg details here, or just leave it with the general idea?

Tape Format Changes

We need to specify the sequence number of the tape in the run, in the tape header file. The file header block specifies whether it is a continuation file or not.

Tapelist File Changes

The lines in the tapelist file should contain the sequence number of the tape in its run, as well as the amount of data written on the tape, and perhaps whether or not the end of tape was reached.