-*- Text -*- Content ======= * Context * Requirements * Nice-to-have's * Non-goals * Open items / discussion points * Problems in wc-1.0 * Possible solutions - Developer sanity - Speed - Cross node type change representation - Flexibility of metadata storage - Transaction duration / memory management - Working copy stability - Transactional updates - Work queue * Prerequisites for a good wc implementation * Modularization * Implementation proposals for - Mapping of svn_wc_entry_t fields to BASE/WORKING - Basic storage mechanics - Metadata schemas - Commit process - Random notes - Code organization - svn_wc.h API * Upgrading old working copies * Implementation plan Context ======= The working copy library has traditionally been a complex piece of machinery and libsvn_wc-1.0 (wc-1.0 hereafter) was more a result of evolution than it was a result of design. This can't be said to be anybody's fault as much as it was unawareness of the developers at the time with the problem(s) inherent to versioning trees instead of files (as was the usual context within CVS). As a result, the WC has been one of the most fragile areas of the Subversion versioning model. The wc is where a large number of issues come together which can be considered separate issues in the remainder of the system, or don't have any effect on the rest of the system at all. The following things come to mind: * Different behaviours required by different use-cases (users) For example: some users want mtime's at checkout time to be the checkout time, some want it to be the historical value at check-in time (and others want different variants). * Different filesystems behave differently, yet Subversion is a cross platform tool and tries to behave the same on all filesystems (timestamp resolution may be an example of this). When considering the wc-1.0 design, one finds that there are a lot of situations where the state of the versioned tree is poorly defined. To clarify the tree state, the wc-ng design splits it into several pieces. A versioned item in the working copy is described by one or more nodes in the following schema. * BASE: Nodes describing repository data in the WC. Each BASE node corresponds to a particular repository URL and revision. Mixed-revision working copies are still common. Pristine file content is kept in the content store. * WORKING: Nodes describing structural changes to BASE. An item in the WC may have a BASE node, a WORKING node, or both. The WORKING-only case arises when an item is added, copied-here, or moved-here on a path that doesn't exist in BASE. * NODE_DATA: Nodes describing layered structural changes to WORKING. An example: suppose a directory is replaced, a directory is copied into it, and a file is added to the subdirectory. BASE and WORKING describe the replacement, NODE_DATA describes the copied subdirectory, and NODE_DATA (again) describes the added file. * ACTUAL: Nodes describing content modifications and annotations. An ACTUAL node contains one or more of the following: - Text-modified flag (for a file) - Properties - Changelist name - Conflict metadata The text-modified flag may be out of sync with the file's real status. The other data is set directly by Subversion operations. Some tree conflicts lead to ACTUAL nodes that don't exist in the working copy (e.g., when a merge tries to edit a file that doesn't exist). These have no corresponding BASE, WORKING or NODE_DATA nodes. ##BHB: For tree, text and property conflicts it would be nice to handle MINE, THEIRS (and OLDER?) as semi-trees too. See this mail thread http://subversion.tigris.org/ds/viewMessage.do?dsForumId=462&dsMessageId=984982 This would provide a clean way to access all information about the conflict origins: $ svn info file@THEIRS Requirements ============ * Developer sanity From this requirement, a number of additional ones follow: - Very explicit tree state management; clear difference between each of the 5 states we may be looking at - It must be "fun" to code wc-ng enhancements * Speed (Note: a trade off may be required for 'checkout' vs 'status' speed) * Cross-node-type working copy changes * Flexibility The model should make it easy to support - central vs local metadata storage - Last modified timestamp behaviours - .svn-less working copy subtrees - different file-changed detection schemes (e.g. full tree scan as in wc-1.0 as well as 'p4 edit') * Graceful (defined) fallback for non-supported operations When a checkout tries to create a symlink on an OS which supports them, on a filesystem which doesn't, we should cope without canceling the complete checkout. Same for marking metadata read-only. * Gracefully handle symlinks in relation to any special-handling of files (don't special-handle symlinks!) * Clear/reparable tree state Other than our current loggy system, I mean here: "there is a command by which the user can restart the command he/she last issued and Subversion will help complete that command", which differs from our loggy system in the way that it will return the working copy to a defined (but to the user unknown) state. * Transactional/ repairable tree state (with which I mean something which achieves the same as our loggy system, but better). * Case sensitive filesystem aware / resilient * Working copy stability; a number of scenario's with switch and update obstructions used to leave the working copy unrecoverable * Client side 'true renames' support where one side can't be committed without the other (relates to issue #876) ###JSS: Perhaps this is obvious... I think that requirement is fine for the user doing the commit. We still need to remember that another user doing the update may not have authz permission to the directory it was renamed into or may have a checkout of a sub-tree and that target directory may not exist. Likewise, the original location might be unavailable too. * Change detection should become entirely internal to libsvn_wc (referring to the fact that libsvn_client currently calls svn_wait_for_timestamps()), even though under 'use-commit-times=yes', this waiting is completely useless. * Last-modified recording as a preparation for solving issue #1256 and as defined in this mail, also linked from the issue: http://svn.haxx.se/dev/archive-2006-10/0193.shtml * Representing "this node is part of a replaced-with-history tree and I'm *not* in the replacement tree" as well as "... and I'm deleted from the replacement tree" [issues #1962 and #2690] Would-be-very-nice-to-have's ============================ * Multiple users with a single working copy (aka shared working copy) * Ending up with an implementation which can use current WCs (without conversion) * Working copies/ metadata storages without local storage of text-bases (other than a few cached ones) Non-goals ========= * Off-line commits * Distributed VC Open items / discussion points ============================== * Files changed during the window "sent as part of commit" to "post commit wc processing"; these are currently explicitly supported. Do we want to keep this support (at the cost of speed)? * Single working copy lock. Should we have one lock which locks the entire working copy, disabling any parallel actions on disjoint parts of the working copy? * Meta data physical read-only marking (as in wc-1.0). Is it still required, or should it become advisory (ie ignore errors on failure)? * Is issue #1599 a real use-case we need to address? (Loosing and regaining authz access with updates in between) Problems in wc-1.0 ================== * There's no way to clear unused parts of the entries cache * The code is littered with path calculations in order to access different parts of the working copy (incl. admin areas) * The code is littered with direct accesses to both wc files and admin area files * It's not always clear at which time log files are being processed (ie transactions are being committed), meaning it's not always clear at which version of a tree one is looking at: the pre or post transformation versions... * There's no support for nested transactions (even though some functions want to start a new transaction, regardless whether one was already started) * It's very hard to determine when an action needs to be written to a transaction or needs to be executed directly * All code assumes local access to admin (meta)data * The transaction system contains non-runnable commands * It's possible to generate combinations of commands, each of which is runnable, but the series isn't * Long if() blocks to sort through all possible states of WORKING, ACTUAL and BASE, without calling it that. * Large if() blocks dealing with the difference between file and directory nodes * Many special-handling if()s for svn:special files * Manipulation of paths, URLs and base-text paths in 1 function * 'Switchedness' of subdirectories has to be derived from the URLs of the parent and the child, but copied nodes also have non-parent-child source URLs... (confusing) * Duplication of data: a 'copied' boolean and a 'copy_source' URL field * Checkouts fail when checking out files of different casing to a case insensitive filesystem * Checkouts fail when marking working copy admin data as read-only is a non-supported FS operation (VFAT or Samba mounts on Linux have this behaviour) * Obstructed updates leave operations half done; in case of a switch, it's not always possible to switch back (because the switch itself may have left now-unversioned items behind) * Directories which have their own children merged into them (which happens when merging a directory-add) won't correctly fold the children into schedule==normal, but instead leave them as schedule==add, resulting in a double commit (through HTTP, other RA layers fold the double add, but that's not the point) [see issue #1962] * transaction files (ie log files) are XML files, requiring correct encoding of characters and other values; given the short expected life-time of a log file and the fact that we're almost completely sure the log file is going to be read by the WC library anyway (no interchange problems), this is a waste of processing time * No strict separation between public and internal APIs: many public APIs also used internally, growing arguments which *should* only matter for internal use * The lock strategy requires writing a file in every directory of a working copy, which severely reduces our performance in several environments. (Windows, NFS). Testing showed that in some cases we used more than 50 seconds on writing 8000 lockfiles before we even started looking what to update. A new lock strategy should reduce the number of writes necessary for locking with depth infinity. Possible solutions ================== Developer sanity ---------------- Strict separation between modules should help keep code focused at one task. Probably some of the required user-specific behaviours can (and should) be hidden behind vtables; for example: setting the file stamp to the commit time, last recorded time or leaving it at the current time should be abstracted from. Access to 'text bases' is another one of these areas: most routines in wc-1.0 don't actually need access to a file (a stream would be fine as well), but since the files are there, availability is assumed. When abstracting all access into streams, the actual administration of the BASE tree can be abstracted from: for all we know the 'tree storage module' may be reading the stream directly off the repository server. [The only module in wc-1.0 which *requires* access to the files is the diff/merge library, because it rewinds to the start of the file during its processing; an operation not supported by streams... and even then, if these routines are passed file handles, they'll be quite happy, meaning they still don't need to know where the text base / source file is...] ###GJS: the APIs should use streams so that we can decompress as the stream is being read. the diff library will need a callback of some kind to perform the rewind, which will effectively just close and reopen the stream. if it rewinds *multiple* times, then we may want to cache the decompressed version of the file. I'll investigate. Given our metadata/base-text storage system, I suspect it will be very easy to cache decompressed copies for a while. ###GJS: a very reasonable strategy is: non-binary files are compressed by default. binaries are stored uncompressed. future improvement: extension-based choices, or some other control In order to keep developers sane, it should be extremely clear at any one time - when operating on a tree - which tree is being operated upon. One way to prevent the lengthy 'if()' blocks currently in wc-1.0, would be to design a dispatch mechanism based on the path-state in WORKING/BASE and the required transformation, dispatching to (small) functions which perform solely that specific task. #####XBC Do please note that this suggests yet another instance of pure polymorphism coded in C. This runs contrary to the developer sanity requirement. ###GJS: agreed with XBC. Speed ----- wc-1.0 assumes the WORKING tree and the ACTUAL tree match, but then goes out of its way to assure they actually do when deemed important. The result is a library which calls stat() a lot more often than need be. One of the possible improvements would be to make wc-ng read all of the ACTUAL state (concentrated in one place, using apr_stat()), keeping it around as long as required, matching it with the WORKING state before operating on either (not only when deemed important!). ###GJS: working copy file counts are unbounded, so we need to be careful about keeping "all" stat results in memory. I'll certainly keep this in mind, however. Working from the ACTUAL tree will also prove to be a step toward clarity regarding the exact tree which is being operated upon. [This suggestion from wc-improvements also applies to wc-ng:] Most operations are I/O bound and have CPU to spare. Consider the virtue of compressed text bases in order to reduce the amount of I/O required. Another idea to reduce I/O is to eliminate atomic-rename-into-place for the metadata part of the working copy: if a file is completely written, store the name of the base-text/prop-text in the entries file, which gets rewritten on most wc-transformations anyway. ###GJS: actually, I believe we *rarely* do full walks over the filesystem multiple times. I doubt we will need to cache stat() information. I think performance will primarily derive from omitting the .svn/ location/walks and opening of multiple files therein, in favor of a single SQLite database open. of course, we'll analyze the situation, but I suspect that we will be in great shape as a natural fallout of our new storage system. Cross node type change representation ------------------------------------- ###GJS: this is not allowed in wc-1, but should be easily possible in wc-ng. the WORKING tree's node kind is different than the BASE tree. no big deal. Flexibility of metadata storage ------------------------------- There are 3 known models for storing metadata as requested by different groups of users: - in-subtree metadata storage (.svn subdir model, as in wc-1.0) ###GJS: euh... aren't we axing this? who has *requested* this? - in-'tree root' metadata storage (working copy central) - detached metadata storage (user-central) - in $HOME/.subversion/ - in arbitrary location (e.g. $HOME is a (slow) NFS mount, and we want the metadata on a local drive, such as /var/...) A solution to implementing each of these behaviours in order to satisfy the wide range of use-cases they solve, would be to define a module interface and implement this interface three times (possibly using vtables). Note that using within-module vtables should be less problematic than our post-1.0 experiences with public vtables (such as the ra-layer vtable): implementation details are allowed to differ between releases (even patch releases). ###GJS: note that we are talking about both metadata AND base-text content. (and yeah, optional and compresses base-texts can be done during this rewrite) Also note that we might be able to share base-text content across working copies if they are all keyed by the MD5 hash into storage directories (under the user-central model) ###GJS: I don't think vtables are needed here. This is simply altering the base location, not a whole new implementation. My plan is to default to the "tree root" model with a .svn subdirectory. If a .svn subdir is not found, then we fall back to looking in the $HOME/.subversion/ directory (some subdir under there). If we *still* don't find it, then some config options will point us to the metadata/base-text location. ###GJS: my plan is to upgrade the working copy if we find a pre-1.7 working copy. all the data will be lifted from the multiple .svn subdirectories, and relocated to the "proper" storage location. This will be a non-reversable upgrade, and will preclude pre-1.7 clients from using that working copy again. Note: because of the "destructive" nature of this upgrade, and the expected duration, we will require the user to perform an explicit action ('svn upgrade') in order to complete the upgrade. However, 1.7 will not be able to *modify* wc-1.0 metadata -- just read it in order to upgrade it to the new storage system. When svn detects an old working copy, then it will error out and request that the user run "svn upgrade" to upgrade their working copy to the new format. The metadata location is determined at one of two points: * checkout time * upgrade time According to the user's config, the metadata will be placed in one of three areas: wcroot: at the root of the working copy in a .svn subdirectory home: in the .subversion/wc/ subdirectory /some/path: stored in the given path All wcroot directories will have a .svn subdirectory. In that directory will be the datastore, or there will be a file that provides two pieces of information: * absolute path to the (centralized) metadata * absolute path of where this wcroot was created With this information, we can link a wcroot to its metadata in the centralized store. If the user has moved the wcroot (the stored path is different from the current/actual path), then Subversion will exit with an error. The user must then ###somehow tell svn that the wc has been copied (duplicate the metadata for the wcroot) or moved (tweak the path stored in the metadata and in the linkage file). Subversion is unable to programmatically determine which operation was used. Note that we use "svn upgrade" as the trigger to *perform* the upgrade. The amount of file opens, parsing, moving, deleting, etc is expected to consume significant amounts of I/O and (thus) cannot simply be done on-the-fly without the user's knowledge and consent. Transaction duration / memory management ---------------------------------------- The current pool-based memory management system is very good at managing memory in a transaction-based processing model. In the wc library, a 'transaction' often spans more than one call into the library. We either need a sane way to handle this kind of situation using pools, or we may need a different memory management strategy in wc-ng. Update (2009-05-10): pool-based management is still being used. we are switching to a "dual pool" system that clarifies the intent of the pools. the result_pool is used for return allocations, and scratch_pool for any temporary allocations. Working copy stability ---------------------- In light of obstructed updates it may not always be desirable to be able to resume the current operation (as currently is the case): in some cases the user may want to abort the operation, in other cases the user may want to resolve the obstruction before re-executing the operation. The solution to this problem could be 'atomic updates': receiving the full working copy transformation, verifying prerequisites, creating replacement files and directories and when all that succeeds, update the working copy. Full workin' copy unit tests: Exactly because the working copy is such an important part of the Subversion experience *and* because of the 'reputation' of wc-1.0, we need a way to ensure wc-ng completely performs according to our expectations. *The* way to ensure we're able to test the most contrived edge-cases is to develop a full unit testing test-suite while developing wc-ng. This will both be a measure to ensure working copy stability as well as developer sanity: in the early stages of the wc-ng develop- ment process, we'll be able to assess how well the design holds up under more difficult 'weather'. ###GJS: agreed. as much as possible, when I (re)implement the old APIs in terms of the new APIs, then I'll write a whitebox test. we'll see how long I keep that up :-P Update (2009-05-10): wc-ng currently passes the entire test suite. Additional tests have been implemented ('entries_tests.py' and other) to try and ensure this continued compatibility. Transactional updates --------------------- .. where 'update' is meant as 'user command', not 'svn update' per se. When applied to files, this can be summarized as: * Receive transformations (update, delete, add) from the server, Work Queue ---------- Certain operations that affect the filesystem require a stateful marker that the operation needs to happen. The best example is when a merge conflict occurs: several "pristine" files need to be placed into the working copy (e.g. somefile.c.r34). Should processing fail after the first file is placed, then we need to "remember" to resume the operation and place the rest of the files. The record of these needed operations will be placed into a "work queue" which is a table recorded in the SQLite database. Much like the original loggy, a working copy will be unusable until these actions are run to completion. Each work item must have the following properties: * order-independent. the work items must be allowed to execute in any sequence. * idempotent. the work item must be able to run an arbitrary number of times. * resumable. whether a previous run completed, or was only partially completed, the work item must be able to complete its operation. * independent. each work item must affect only one node in the logical trees. it can apply to any/all of BASE/WORKING/ACTUAL, but it must apply to a single logical node. * complete. a work item must represent a complete operation which takes the WC from one stable state to another. thus, a work item cannot be used to "return the wc to a stable state" (the operation that made it unstable should be included in the work item). The goal here is to reduce interactions across work items. Each must be completely self-sufficient and resumable. The wc_db API will provide a low-level framework for adding, fetching, and completing these work items. Each work item will be described by a skel, to be interpreted by higher levels. ### the "independent" requirement is subject to discussion. It may be possible to have a work item that touches multiple nodes. As long as it can definitely place those nodes into a specific state, then it might be okay to operate on many. Prerequisites for a good wc implementation ========================================== These prerequisites are to be addressed, either as definitions in this document, or elsewhere in the subversion (source) tree: * Well defined behaviour for cross-node type updates/merges/.. (tree conflicts in particular) * Well defined behaviour for special file handling * Well defined behaviour for operations on locally missing items (see issue #1082) * Well defined change detection scheme for each of the different last-modified handling strategies * No special handling of symlinks: they are first class versioned objects * Well defined behaviour for property changes on updates/merges/... (this is a problem which may resemble tree conflicts!), including 'svn:' special properties * File name manipulation routines (availability) * File name comparison routines (!) (availability; which compensate for the different ways Unicode characters can be represented [re: NFC/NFD Unicode issue]) ###JSS: Talking with ehu on IRC when I asked him about how to handle this issue: "if we accept that some repositories will be unusable with wc-ng, then we can standardize anything that comes in from the server as well as the directory side into the same encoding. we'd be writing files with the standardized encoding." The rest of this conversation centered around the fact that either APR or the OS will convert the filename to the correct form for the filesystem when doing the stat() call. Note, ehu says: "(we'll need to retain the filename we got from the server though: we'll need it to describe the file through the editor interface: the server still allows all encodings.)" * URL manipulation routines (availability) * URL comparison routines (availability; which compensate for different ways the same URL can be encoded; see issue #2490) * Modularization * Agree on a UI to pull in other parts of the same repository (NOT svn:externals) [relates to issue #1167] #####XBC I submit this is a server-side feature that the client (i.e. the WC library) should not know about. * Agree on behaviour for update on moved items (relates to issue #1736) * Case-sensitivity detection code to probe working copy filesystem Implementation proposals ======================== Classification of svn_wc_entry_t fields to BASE/WORKING ------------------------------------------------------- [Note: This section is mainly to clarify the difference between the BASE and WORKING trees, it's not here to mean that we actually need all these fields in wc-ng!] Here are the mappings of all fields from svn_wc_entry_t to the BASE and WORKING trees: +-------------------------------+------+---------+ | svn_wc_entry_t | BASE | WORKING | +-------------------------------+------+---------+ | name | x | x (1)| | revision | x | x (2)| | url | x | x (2)| | repos | x | x (3)| | uuid | x | x (3)| | kind | x | x | | absent | x | | | copyfrom_url | | x | | copyfrom_rev | | x | | conflict_old | | x | | conflict_new | | x | | conflict_wrk | | x | | prejfile | | x | | text_time | | = | | prop_time | | = | | checksum | x | x (2)| | cmt_rev | x | x (2)| | cmt_date | x | x (2)| | cmt_author | x | x (2)| | lock_token | x(6)| | | lock_owner | x | | | lock_comment | x | | | lock_creation_date | x | | | has_props | x | x (4)| | has_prop_mods | | = | | cachable_props | x(5)| x (4)| | present_props | x | x (4)| | changelist | | x | | working_size | | = | | keep_local | | = | | depth | x | x | | schedule | | | | copied | | | | deleted | | | | incomplete | | | +-------------------------------+------+---------+ (1) if this one differs from BASE, it must point to the source of a rename (2) for an add-with-history (3) or can we assume single-repository working copies? (4) can differ from BASE for add-with-history (5) why is this a field at all; can't the WC code know? (6) locks apply to in-repository paths, hence BASE The fields marked with '=' are implementation details of internal detection mechanisms, which means they don't belong in the public interface. Fields with no check are to become obsolete. 'schedule', 'copied' and 'deleted' can be deducded from the difference between the BASE and WORKING or WORKING and ACTUAL trees. 'incomplete' should become obsolete when the goal of 'atomic updates' can be realised, in which case the tree can't be in an incomplete yet locked state. This would also invalidate issue #1879. Basic Storage Mechanics ----------------------- All metadata will be stored into a single SQLite database. This includes all of the "entry" fields *and* all of the properties attached to the files/directories. SQLite transactions will be used rather than the "loggy" mechanics of wc-1.0. ###GJS: note that atomicity across the sqlite database and the content of the ACTUAL tree is freakin' difficult. idea to test: metadata says "not sure of ACTUAL", and when ops complete successfully, then we clear the flag. during any future operation, if the flag is present, then we approach the ACTUAL with extreme prejudice. also note that we can batch clearing of the flags as an optimistic efficiency approach (since if we batch 100 and the last fails, then the other 99 will be slower until the wc-ng determines the ACTUAL is in fine shape and clears the flag for future operations). ###GJS: be wary of sqlite commit performance (based on some of my prior experience with it). must have timing/debugging around the commit operations. may need to use various transaction isolations and/or batching of commits to get proper performance. thus, profile output capability is mandatory to determine if we have issues, and where they occur. ###JSS: I don't see how transactions by themselves can replace loggy. Right now, if you abort something like 'svn update' or 'svn checkout', loggy has recorded all the files to be downloaded, and will pick up where it left off. We did this as an optimization to prevent re-downloading a potentially large amount of data again. Seems like we still need to provide that capability. ###GJS: sqlite transactions replace the atomicity that loggy was originally designed for. it sounds like loggy is also be used as a work queue, and that is easily handled in sqlite. Base text data will be stored in a multi-level directory structure, keyed/named by the checksum (MD5 or SHA1) of the file. The database will record appropriate mappings, content/compression types, and refcounts for the base text files (for the shared case). We will use a single level of directories: TEXT_BASE/7c/7ca344... With 100k files spread across all of a user's working copies, that will put 390 files into each subdirectory, which is quite fine. If the user grows to a million files, then 3900 per subdir is still reasonable. Two levels would effectively mean one file per subdir in typical situations, which is a lot of disk overhead. When the metadata is recorded in a central area (rather than the WC root), then it is possible for the metadata and the base files to become out of date with respect to all the working copies on the system. We will revamp "svn cleanup" to re-tally the base text reference counts, eliminate unreferenced bases, verify that the working copies are still present, ensure the metadata <-> WC integrity, deal with moves of metadata from central -> wc-root (can happen if somebody rm -rf's the wc, then does a checkout and wants the metadata at the wc-root (this time)), and other consistency checks. Metadata Schemas ---------------- see libsvn_wc/wc-metatdata.sql3 The table below describes, in English, the various combinations of "presence" values as the occur in the BASE_NODE and WORKING_NODE tables. BASE_NODE WORKING_NODE DESCRIPTION normal Node has been checked out from the repository normally. absent Server has marked the node as "absent", meaning the user does not have authorization to view the content. excluded The node has been (locally) marked as excluded from the working copy. not-present The node is not present at its current revision. The parent directory has a different revision which states the node *is* present. This state is usually reached by locally deleting a file and committing it. Later, when an update is run, the directory will be bumped to a revision that does not contain the file, and this not-present node will be cleared. incomplete The node is known, but the node's information has not been downloaded (yet) from the server. This typically occurs from an interrupted checkout. The parent directory was added, specifying all the children, but the checkout was stopped before fetching the child. base-deleted * Not allowed. This presence is only valid for the WORKING_NODE table. absent Not allowed. The name exists on the server and cannot be modified in any way. incomplete An update "under" some local changes was terminated before fetching information on this node. normal This node has been locally-added through a simple add, a copy, or a move (other data needs to be examined to determine what operation brought the node here). normal normal The underlying BASE node has been deleted, and a new node has been added in its place (this is a "Replace"). excluded normal A node was excluded from the checkout or update, and we have locally-added a new node to *replace* it. not-present normal A node is no longer present in the BASE tree due to mixed-revision working copy concerns. This is an addition (not a replace) of a new node, in the same location as a node that deleted (and committed) at some point in history. * absent Not allowed. This would imply that the server has prevented our access, but this is a local, uncommitted change. The server cannot block the node. (see Note 1) * excluded An add-with-history or a move has been performed, and this node has been excluded from the working copy. Note: plain adds cannot have an excluded node -- we'd just not add the node. Further note: the root of the copy/move cannot be excluded since we need the source information. The root may be depth==empty, however. * not-present This node has been locally-deleted. This can only occur for a child of a copied or moved subtree (for a plain add, we simply revert the add; and must be a child, or we'd just revert the whole copy or move operation). * incomplete This node is known, but the information is missing. A copy, move, or deletion has been interrupted, leaving a directory with known children, but lacking their state. (see Note 2) normal base-deleted The BASE node has been locally-deleted. excluded base-deleted The node was excluded from the working copy, but has been locally-deleted. (see Note 3) not-present base-deleted Not allowed. There is nothing to delete. The not-present is a tool to represent mixed-rev working copies; there is (logically) no node to delete. "" means there is no row in the given table (so no presence value) "" means "any value" "*" means " or " Note 1: this implies you cannot copy/move a working copy tree that has absent nodes in it. If that were made possible, then we may (instead) want to model this as a copy/move followed by a local-delete of the absent node(s). Note 2: this will probably only apply to a repository-to-wc copy. For wc-to-wc copies/moves, we will probably transact the entire operation so that a child will never be incomplete. Note 3: this may be possible, though we may need more state to pull it off (eg. what revision of the node is "not there" is will be deleted? need the rev for out-of-date checks) Commit Process -------------- Committing is essentially a review of all the rows in the WORKING_NODE and ACTUAL_NODE tables, and sending appropriate instructions to the server. After the commit, those rows are removed and the data "collapsed" down into the BASE_NODE. For example, if a copy has been performed, then WORKING_NODE contains data about the copy, and the new BASE_NODE will "become" that WORKING_NODE. If the user additionally modifies some properties (stored in ACTUAL_NODE), then those will also fall down into the post-commit BASE_NODE row. copy_tests 24 sets up a specific scenario that breaks a lot of the code in libsvn_wc today (Sep 28, 2009). While the commit is being processed, the database is temporarily placed into a state which does have not have proper integrity. This section is an attempt to document rules for how nodes are to be treated during the commit process. These rules focus on the effects upon an individual node, independent of what happens to any child nodes on the theory that children can be configured as a mixed-ref working copy with appropriate operations applied. Note that the decision process is independent of the children, but the children *will* be affected by the commit of the parent node. The commit MUST operate in a top-down fashion, however, since (for example) it is impossible to model a copied parent and an unmodified (committed) BASE child. There is a large operational difference between directories and files/symlinks, so we'll divide the discussion along that line. Note that we only consider the kind of the WORKING node; the kind of the BASE will simply be replaced by the new node. If the BASE used to be a directory, then its (obsolete) children must be removed from BASE_NODE during this commit proces. NOTE: if either the BASE_NODE or the WORKING_NODE has an "incomplete" presence, then it CAN NOT be committed. It means we are missing information that may be required to properly commit a change to that node. ### hmm. during the commit, we create incomplete BASE nodes (see ### below). so this is more of a statement of the starting condition. FILES AND SYMLINKS working-presence: normal base-presence: * The WORKING and ACTUAL data is collapsed down into BASE_NODE, with the new revision. working-presence: excluded base-presence: * A new, excluded BASE_NODE is constructed, and the WORKING_NODE is removed. Any BASE_NODE rows which appear to be descendents of this (used-to-be-directory) node are removed. There should be no descendents in the WORKING_NODE table. ### what information do we keep for excluded nodes? ### note: at this point, there is no user command to exclude ### files/symlinks. but we will be able to at some point... working-presence: not-present base-presence: * NOTE: this situation should never be seen, since the node's parent should have been committed first, which handles this node as part of the child processing (see resolution for directories). ### in short, this work-presence can only occur for a deleted ### *subroot*. and we cannot commit *just* this node. must commit ### the root of any copy/move operation. working-presence: base-deleted base-presence: * The WORKING_NODE and BASE_NODE rows are removed. ### if base-presence == excluded, there are some concerns: ### -- what information do we keep for base=excluded nodes? ### -- the base-deleted node would (at least) need to retain the ### revision in order to mark it for deletion. ### -- note: at this point, there is no user command to exclude ### files/symlinks. but we will be able to at some point... actual: row is available working-presence: base-presence: normal ACTUAL_NODE.properties are folded into the BASE_NODE and the revision is bumped. Note: there is no data in ACTUAL_NODE other than properties. Conflict information exists, but that must be cleared before committing is possible. DIRECTORIES Whenever a directory is bumped to a new revision, the new set of children is provided. This is required, in order to maintain the proper integrity. Example: if a child is to be added in this new revision, but a failure happened between the directory-commit and the add-child processes, then there would be no record of the added child. The directory would not know it was missing a child and would report "at revision R" to the server, implying that the child is present. working-presence: normal base-presence: * The WORKING and ACTUAL data is collapsed down into BASE_NODE, with the new revision. The depth status of the WORKING node is carried over to the BASE node. Children in the WORKING_NODE table should align with that depth value, and the commit will iterate over each available child row. Its status will be examined, and specific action taken: normal An incomplete node should be added for this child. This node will become its own add/copy/move root, and will be handled as a separate action (via recursion over the children). excluded An incomplete node should be added for this child. No action taken. This node will become an excluded BASE node when it is handled as a separate action (via recursion). not-present This row in WORKING_NODE is removed, along with descendent nodes. The directory will not list this node in its (new) set of children. Any BASE_NODE row at this path is also removed, along with any descendent nodes. base-deleted No action taken. This node will be removed when it is handled as a separate action (via recursion). working-presence: excluded base-presence: * A new, excluded BASE_NODE is constructed, and the WORKING_NODE is removed. Any BASE_NODE rows which appear to be descendents of this (used-to-be-directory) node are removed. There should be no descendents in the WORKING_NODE table. ### what information do we keep for excluded nodes? working-presence: not-present base-presence: * NOTE: this situation should never be seen, since the node's parent should have been committed first, which handles this node as part of the child processing (see resolution). working-presence: base-deleted base-presence: * The WORKING_NODE and BASE_NODE rows are removed. actual: row is available working-presence: base-presence: normal ACTUAL_NODE.properties are folded into the BASE_NODE and the revision is bumped. Note: there is no data in ACTUAL_NODE other than properties. Conflict information exists, but that must be cleared before committing is possible. Random Notes ------------ ### break down all modification operations to things that operate on a small/fixed set of rows. if a large sequence of operations fails, then it can leave the system in reparable state, since most were performed. note that ACTUAL can change at any time, thus all mods should be able to compensate for ACTUAL being something unexpected. thus, the transformative operations should be able to fail in such a way as to leave ACTUAL pretty bunged up. ### probably want to special-case the checksum and BASETEXT entry for the "empty file" Code Organization ----------------- libsvn_wc/wc_db.h (symbols: svn_wc__db_*) Storage subsystem for the WC metadata/base-text information. This is a private API, and the rest of the WC will be rebuilt on top of this. This code deals with storage, and transactional modifications of the data. Note: this is a random-access, low-level API. Editors will be built on top of this layer. libsvn_wc/workqueue.h (symbols: svn_wc__wq_*) The "work queue" is a subsystem to replace the old "loggy" subsystem. It will perform (primarily) filesystem operations in a transactional way. svn_wc.h API ------------ Note that we also have an opportunity to revamp the WC API. Things like access batons will definitely disappear, but there will most likely be great opportunities for other design changes. Note that removing access batons (and other API changes) will ripple up until libsvn_client, and may even have an affect on *its* API. ### the form of a new API is unknown/TBD. We are going to add svn_wc_context_t to be created once per process, and passed to all svn_wc functions. This will replace the (often confusing) use of access batons. Implementation note: this context will hold an svn_wc__db_t handle, and a pointer to the process's svn_config_t object. Upgrading old working copies ============================ When WC-NG finds a working copy which is pre-wc-ng, it will quit, prompting the user to run 'svn upgrade' to upgrade the working copy to a wc-ng state. The reason for not upgrading on-the-fly is two-fold: * We anticipate this process to be irreversible, so we want to ensure the user wants to upgrade (no silent upgrading/breakage). * The upgrade may be I/O and computationally intensive, and keeping with the principle of least surprise, we want to ensure that the upgrade is done intentionally, when the user expects it. Here's how we plan on implementing 'svn upgrade', so that it maintains consistency across the working copy atomically. We have two requirements: do the upgrade completely, and *don't* leave the working copy in an unusable state if the upgrade fails. The steps for upgrading are: 1) Create the new wc.db as invisible.db in the working copy root 2) Upgrade the current directory into invisible.db 3) Drop a flag file into the directory to signal an "in process" upgrade 4) Recurse on step 2 for each of the subdirectories 5) Move invisible.db to wc.db in the working copy root 6) Recursively remove each of the .svn subdirs for each wc subdir Note that nowhere do we attempt to run or upgrade old logs. This is intentional. In order to simplify the development and maintenance burden, we intend to bail when the upgrade process encounters a working copy with un-run logs. In this state, it will be up to the user to run 'svn cleanup' with the prior version of Subversion to ensure the working copy is in an upgradable state. Failing that, the user can always do a fresh checkout. The atomic step is step 5. Should the upgrade process be interrupted prior to Step 5, the working copy will still be usable by a pre-wc-ng client, but will just have extra stuff in the .svn directories, namely invisible.db in the root, and the various flag files everywhere. Should the upgrade get interrupted *after* Step 5 (but before all the .svn directories are removed), the .svn directories will show up as unversioned directories. Not the ideal, but not terribly bad, either. ### perhaps we should have a way of cleaning up all the .svn dirs from an interrupted upgrade? ### what if somebody attempts to use an old Subversion on a working copy with a .svn which hasn't been harvested yet? it should succeed, but may leave a discrepancy between the wc-ng database and the .svn metadata, since old Subversions don't know to recurse up the tree. could we do something to the working copy to render it unusable by older subversions, but without too much pain? how about simply bumping the format number in entries.c? a properly chosen format number would make older Subversions complain, but also make Subversion 1.7 prompt for an upgrade. hmm..... When an upgrade is restarted, invisible.db will just be blown away and recreated from scratch, since already-upgraded directories could have been modified between invocations of 'svn upgrade'. Implementation Plan =================== The following are tests which need to be accomplished for WC-NG. There isn't a strict ordering here, but rather a possible plan. There may be dependencies between some items, but that is left as an exercise for the reader. * Pristine file management * Properties management * Tree management (BASE v. WORKING v. ACTUAL for APIs and storage) * Journaled actions * Finding/using the correct admin area * Upgrading - Including multiple heterogenous admin areas * Move entries into SQLite * Relocating datastore in useful ways Afterwards, we'll need: * A second pass at the WC code to find/fix patterns and solutions. * Revamp of WC API, to propagate up into libsvn_client. * Reexamine any client/wc interactions, and look for final cleanups. Near-Term Plan -------------- Note: we originally envisioned the "ordering" below. In practice, however, we have been attacking the overall problem from multiple angles. Typically, we are finding conceptual/API bottlenecks that make it hard to accomplish a number of other tasks. We solve the bottleneck, and move on to solving the higher-level problems. It is a continual "evolutionary" process. Some temporary APIs are being introduced to help bridge the conceptual gap between wc-1 and wc-ng. These should disappear by release time, but serve to mitigate code disruption and potential for error. 1. convert entries.c to use sqlite directly. migrate 'entries' file during this step. the sqlite file will be in-memory if we are not allowed to auto-upgrade the WC; otherwise, we'll write the sqlite database into .svn/ note: the presence of 'wc.db' (or whatever its name) will indicate a minimum format level. the user field in the database contains the schema version which is our further format-level descriptor value. [ this has been largely done. ] 2. convert entries.c to use wc_db. shift the sqlite code into wc_db. note: this is a separate step from 1. there is a paradigm shift between how entries.c works and wc_db works. we want to ignore that in Step 1, and then handle it in this Step. note: put wc_db handle into lock->shared and share the handle across all directories/batons. [ because of the borken way the upper layers use the entries API, using the wc_db APIs to write entries proves difficult, since we violate all kinds of constraints. ] 3. convert props.c to use wc_db. migrate props to db simultaneously. [ this is currently in-process. ] 4. implement 'svn cleanup' as an upgrade path from old-style working copies to wc-ng [ done, but work will continue as the wc-ng format continues to evolve. ] 5. incremental shift of pristines from N files into pristine db. note: we could continue to leave .revert-base while we migrate the primary base into the pristine dataset. 6. shift libsvn_wc from using entries.h to using wc_db.h. note: since entries.h is "merely" a wrapper for wc_db.h, this will allow the libsvn_wc to start using the new wc_db APIs wherever it is easy/possible. goal: all libsvn_wc code uses wc_db.h, and entries.h exists solely to support old backwards-compat code. [ this isn't quite a discreet task, but is happening gradually as we work through the libsvn_wc API. ] 7. centralize the metadata and pristines note: this will also involve merging datastores 8. replace loggy with sqlite-based work journal Endgame ------- As WC-NG development has progress, and many of the above milestones have been met, we've identified the following milestones leading to the completion of wc-ng development (and hence 1.7). They are not all necessarily serially dependent, but some dependencies do exist. 1. Move properties into wc.db. 2. Convert loggy actions into work queue actions. 3. Move pristines into a SHA-1 based store. 4. Consolidate metadata into the centralized system. 5. Test, tweak and release! The above items are milestones. There are a number of work items that need to be completed in/around the milestones. The progress of this work can be roughly measured by the tools/dev/wc-ng/count-progress.py script: * remove use of svn_wc_adm_access_t and svn_wc_entry_t * review/revamp the requirements, definitions, and use of the svn_wc__node_* and svn_wc__db_temp_* functions