ceph/branches/sage/mds/TODO

some smallish projects:

- crush rewrite in C
 - generalize any memory management etc. to allow use in kernel and userspace
- userspace crush tools
 - xml import/export?  
 - ?

- pg monitor service
 - to support statfs?
 - general pg health
 - some sort of (throttled) osd status reporting
 - dynamic pg creation (eventually!)

- SimpleMessenger
 - clean up/merge Messenger/Dispatcher interfaces
 - auto close idle connections
 - delivery ack and buffering, and then reconnect
 - take a look at RDS?  http://oss.oracle.com/projects/rds/

- generalize monitor client?
 - throttle message resend attempts

- ENOSPC on client, OSD




code cleanup
- endian portability
- word size
  - clean up all encoded structures

general kernel planning
- soft consistency on (kernel) lookup?
- accurate reconstruction of (syscall) path?

software raid layer for EBOFS?
- actually, we just need software raid _awareness_ in the allocator, so
  that we can write only full stripes, without fear of clobbering things on
  failure.  then use MD or similar layer provided by kernel.


sage doc
- mdsmonitor beacon semantics
- cache expiration, cache invariants
 - including dual expire states, transition, vs subtree grouping of expire messages
- recovery states, implicit barrier are rejoin
- journal content
 - importmaps and up:resolve
- metablob version semantics



sage mds

- the split/merge plan:

- hmm, should we move ESubtreeMap out of the journal?  
  that would avoid all the icky weirdness in shutdown, with periodic logging, etc.

- extend/clean up filepath to allow paths relative to an ino
  - fix path_traverse
  - fix reconnect/rejoin open file weirdness

- stray reintegration
- stray purge on shutdown
  - need to export stray crap to another mds..
- verify stray is empty on shutdown

- consistency points/snapshots
  - dentry versions vs dirfrags...

- more testing of failures + thrashing.
  - is export prep dir open deadlock properly fixed by forge_replica_dir()?
- failures during recovery stages (resolve, rejoin)... make sure rejoin still works!

- detect and deal with client failure
  - failure during reconnect vs clientmap.  although probalby the whole thing needs a larger overhaul...

- inode.max_size
- inode.allocated_size

- real chdir (directory "open")
  - relative metadata ops

- osd needs a set_floor_and_read op for safe failover/STOGITH-like semantics.

- EMetablob should return 'expired' if they have higher versions (and are thus described by a newer journal entry)

- fix rmdir empty exported dirfrag race
  - export all frags <= 1 item?  then we ensure freezing before empty, avoiding any last unlink + export vs rmdir race.
  - how to know full dir size (when trimming)?
    - put frag size/mtime in fragmap in inode?  we will need that anyway for stat on dirs
      - will need to make inode discover/import_decode smart about dirfrag auth
    - or, only put frag size/mtime in inode when frag is closed.  otherwise, soft (journaled) state, possibly on another mds.
  - need to move state from replicas to auth.  simplelock doesn't currently support that.
    - ScatterLock or something?  hrm.

- FIXME how to journal root and stray inode content? 
  - in particular, i care about dirfragtree.. get it on rejoin?
  - and dir sizes, if i add that... also on rejoin?



osdmon
- allow fresh replacement osds.  add osd_created in osdmap, probably
- monitor needs to monitor some osds...
- monitor pg states, notify on out?
- watch osd utilization; adjust overload in cluster map

journaler
- fix up for large events (e.g. imports)
- use set_floor_and_read for safe takeover from possibly-not-quite-dead otherguy.
- should we pad with zeros to avoid splitting individual entries?
  - make it a g_conf flag?
  - have to fix reader to skip over zeros (either <4 bytes for size, or zeroed sizes)
- need to truncate at detected (valid) write_pos to clear out any other partial trailing writes


crush
- xml import/export?
- crush tools


rados snapshots
- integrate revisions into ObjectCacher
- clean up oid.rev vs op.rev in osd+osdc

- attr.crev is rev we were created in.
- oid.rev=0 is "live".  defined for attr.crev <= rev.
- otherwise, defined for attr.crev <= rev < oid.rev  (i.e. oid.rev is upper bound, non-inclusive.)

- write|delete is tagged with op.rev
  - if attr.crev < op.rev
    - we clone to oid.rev=rev (clone keeps old crev)
    - change live attr.crev=rev.
  - apply update
- read is tagged with op.rev
  - if 0, we read from 0 (if it exists).
  - otherwise we choose object rev based on op.rev vs oid.rev, and then verifying attr.crev <= op.rev.

- how to get usage feedback to monitor?

- clean up mds caps release in exporter
- figure out client failure modes
- add connection retry.


objecter
- transaction prepare/commit
- read+floor_lockout

osd/rados
- transaction prepare/commit
  - rollback
  - rollback logging (to fix slow prepare vs rollback race)
- read+floor_lockout for clean STOGITH-like/fencing semantics after failover.

- consider implications of nvram writeahead logs
- clean shutdown?
- pgmonitor should supplement failure detection

- flag missing log entries on crash recovery  --> WRNOOP? or WRLOST?

- efficiently replicate clone() objects
- fix heartbeat wrt new replication
- mark residual pgs obsolete  ???
- rdlocks
- optimize remove wrt recovery pushes
- report crashed pgs?

messenger
- fix messenger shutdown.. we shouldn't delete messenger, since the caller may be referencing it, etc.

simplemessenger
- close idle connections
- buffer sent messages until a receive is acknowledged (handshake!)
  - retry, timeout on connection or transmission failure
- exponential backoff on monitor resend attempts (actually, this should go outside the messenger!)

objectcacher
- ocacher caps transitions vs locks
- test read locks

reliability
- heartbeat vs ping?
- osdmonitor, filter

ebofs
- allow holes

- verify proper behavior of conflicting/overlapping reads of clones
- combine inodes and/or cnodes into same blocks
- allow btree sets instead of maps
- eliminate nodepools
- nonblocking write on missing onodes?
- fix bug in node rotation on insert (and reenable)
- fix NEAR_LAST_FWD (?)

- awareness of underlying software/hardware raid in allocator so that we
  write full stripes _only_.
  - hmm, that's basically just a large block size.

- rewrite the btree code!
  - multithreaded
  - eliminate nodepools
  - allow btree sets
  - allow arbitrary embedded data?
  - allow arbitrary btrees
  - allow root node(s?) to be embedded in onode, or whereever.
  - keys and values can be uniform (fixed-size) or non-uniform.  
    - fixed size (if any) is a value in the btree struct.  
      - negative indicates bytes of length value?  (1 -> 255bytes, 2 -> 65535 bytes, etc.?)
    - non-uniform records preceeded by length.  
    - keys sorted via a comparator defined in btree root.  
      - lexicographically, by default.

- goal
  - object btree key->value payload, not just a data blob payload.
  - better threading behavior.
    - with transactional goodness!

- onode
  - object attributes.. as a btree?
  - blob stream
  - map stream.
    - allow blob values.

  - 



remaining hard problems
- how to cope with file size changes and read/write sharing


crush
- more efficient failure when all/too many osds are down
- allow forcefeed for more complicated rule structures.  (e.g. make force_stack a list< set<int> >)


mds
- distributed client management
- chdir (directory opens!)
- rewrite logstream
  - clean up
  - be smart about rados ack vs reread
  - log locking?  root log object
  - trimming, rotation

- efficient stat for single writers
- lstat vs stat
- add FILE_CAP_EXTEND capability bit
- only share osdmap updates with clients holding capabilities
- delayed replica caps release... we need to set a timer event? (and cancel it when appropriate?)
- finish hard links!
 - reclaim danglers from inode file on discover...
 - fix rename wrt hard links
- interactive hash/unhash interface
- test hashed readdir
- make logstream.flush align itself to stripes

- carefully define/document frozen wrt dir_auth vs hashing



client
- fstat
- mixed lazy and non-lazy io will clobber each others' caps in the buffer cache.. how to isolate..
- test client caps migration w/ mds exports
- some heuristic behavior to consolidate caps to inode auth?







why qsync could be wrong (for very strict POSIX) : varying mds -> client message transit or processing times.
- mds -> 1,2 : qsync
- client1 writes at byte 100
- client1 -> mds : qsync reply (size=100)
- client1 writes at byte 300
- client1 -> client2 (outside channel)
- client2 writes at byte 200
- client2 -> mds : qsync reply (size=200)
-> stat results in size 200, even though at no single point in time was the max size 500.
-> for correct result, need to _stop_ client writers while gathering metadata.


SAGE:

- string table?

- hard links
 - fix MExportAck and others to use dir+dentry, not inode
   (otherwise this all breaks with hard links.. altho it probably needs reworking already!)

- do real permission checks?






ISSUES


- discover
 - soft: authority selectively repicates, or sets a 'forward' flag in reply
 - hard: authority always replicates (eg. discover for export)
 - forward flag (see soft)
 - error flag   (if file not found, etc.)
 - [what was i talking about?] make sure waiters are properly triggered, either upon dir_rep update, or (empty!) discover reply



DOCUMENT
- cache, distributed cache structure and invariants
- export process
- hash/unhash process


TEST
- hashing
 - test hash/unhash operation
 - hash+export: encode list of replicated dir inodes so they can be discovered before import is procesed.
 - test nauthitems (wrt hashing?)


IMPLEMENT

- smarter balancing
  - popularity calculation and management is inconsistent/wrong.
  - does it work?

- dump active config in run output somewhere