Towards replacing rpm + yum (0.1): - drop the filelists from the main package set file, split out to a secondary file. for package sets that depend on other package sets, we need to be able to generate properties with owning packages that are in another set. this way, a package that requires a file, will look up the provides in the set and find the package that owns it and then try mark that for update. - installer part: - pre install check; check that dirs can be created (no files where want to create dirs), move config files according to file flags. (.rpmnew etc) - store rpm headers for installed packages. - implement rpm uninstall and update. - triggers? just say no? - rpm seems to consider glibc > 2.6.90 to mean greater than 2.6.90-anything. That is, a comparison that doesn't mention the release field, shouldn't regard the release field of pkgs it compares against. glibc-common-2.6.90 has conflicts: glibc < 2.6.90, glibc > 2.6.90 since rpm doesn't let you do glibc != 2.6.90, and requires: glibc = 2.6.90 will always pull in glibc. But even with a != relation, would glibc-2.6.90-16 be equal to 2.6.90? glibc 2.7.90-8 dropped it in favor of requires = 2.7.90-8 (#225806). - signed packages - space calculation before transaction, but ideally, do a number of smaller transactions. - pre-link changing binaries and libs on disk screwing up checksum? - pipelined download and install; topo-sort packages in update set, pick one with all deps in the current set, add it to the current set and satisfy deps against update set => result: minimal update transaction. Queue download and install/update transaction for the packages in the minimal set, start over. This also makes the installation phase much more interruptible, basically just stop after a sub-transaction finishes. As we keep the update set around as a target, we can restart if needed. Probably don't need to, can just do a new update. During a sub-transaction we should keep the target set (i.e. the current set to be) around as a lock file (system.rzdb.lock or so, see git) so that razor updates are prevented if the systems crashes during an update. - implement depsolving between multiple package sets by creating an iterator that has a sorted list of all installed pkgs from all sets, all installed requires from all sets, all installed provides from all sets etc. could be a list of tuples (pkgs index, set index). should simplify even the two-set depsolving a bit since we can pretend there's just one set. this should also be useful for the 'overlay set' idea where the system set is actually made up of a number of sets, but typically a read-only set from a read-only fs and a read-write set from a r/w fs. - locking: we use advisory file locking on the system set (/var/lib/razor/system.rzdb) to indicate a transaction is in progress. The locking algorithm is as follows: 1. obtain advisory lock on system set. if this is already taken, we know that a process is actively modifying the system set and we have to wait. there's a fcntl that lets you block for the lock to go away. 2. if a system-next.rzdb file already exists an earlier razor process was interrupted or crashed and we may want to clean that up. the system-next.rzdb file will record what the previous instance was trying to do and we can just replay that to clean up. 3. create the new package set whichever way and write it to system-next.rzdb, then start installing/removing rpms. 4. When the update is complete, rename system-next.rzdb to system.rzdb and remove the advisory lock. we should probably introduce a new object that encapsulates this sequence, the filename conventions, rpm cache, e.g. struct razor_image, with operations such as #define RAZOR_IMAGE_READ 0x01 #define RAZOR_IMAGE_WRITE 0x02 struct razor_image * razor_image_open(const char *root, unsigned int flags); int razor_image_begin_transaction(struct razor_image *image, struct razor_set *target); int razor_image_finish_transaction(struct razor_image *image); the transaction pipelineing described above sits on top of this, since each step there needs to complete a full transaction that writes out a new package set. for overlay package sets we could do something like struct razor_image * razor_image_open_with_base(const char *root, unsigned int flags, struct razor_image *base); where base specifies the r/o package set it's layered on. this allows for stacking several layers of images. Package set file format items: - nail down byte-order of repo file. - version the sections in the file, put the element size in the header so we can add stuff to elements in a backwards compatible way. maybe not necessary, we can just add sections that augment the sections we want to add to (similar to how rpm has add versioned deps). Misc ideas: - keep history of installed packages/journal of package transaction, so we can roll back to yesterday, or see what got installed in the latest yum update. - use hash table for package and property lists so we only store unique lists (like for string pool). - use existing, running system as repo; eg razor update razor://other-box.local evince to pull eg the latest evince and dependencies from another box. We should be able to regenerate a rzr pkg from the system so we can reuse the signature from the originating repo. - Ok, maybe the fastest package set merge method in the end is to use the razor_importer, but use a hash table for the properties. This way we can assign them unique IDs immediately (like tokenizing strings). - test suite should be easy, just keep .rzdb files around and test different type of upgrades that way (obsoletes, conflicts, file conflicts, file/dir problems etc). Or maybe just keep a simple file format ad use a custom importer to create the .rzdb files. - overlay package sets? mount a read-only /usr over nfs or from the virt-host and have a local package set overlaid over the read-only one. shouldn't need new features in the core package set data structure, but should be just conventions on top. we have the base package set from the r/o system, the overlay set from the local system and we can have an effective package set which is the merge of everything from the overlay into the base set. the effective set is easy to compute and we could do it on the fly or cache it. or maybe the effective set is the on-disk representation and the overlay can be computed when needed, we just keep a link back to the base. - incremental rawhide repo updates? instead of downloading 10MB zipped repo every time, download a diff rzdb? Should be pretty small, especially if we don't have file checksums in metadata. Filenames and properties are for the most part already present, typically just a version bump plus maybe tweaking a couple requires. The upstream repo can store multiple incremental updates in one big file and provide an index file that maps updates for a given date (we should use repo-file checksums though) to a range in the file: Download the index file, search for a match for your latest rawhide.rzdb file, download range of updates that brings it up to date. - use hash tables for dirs when importing files to avoid qsorting all files in rawhide. Bugs: - eliminate duplicate entries in package property lists.