~bzr-pqm/bzr/bzr.dev

===========
Inventories
===========

.. contents::

Overview
========

Inventories provide an abstraction for talking about the shape of a tree.
Generally only tree object implementors should be concerned about entire
inventory objects and their implementation. Other common exceptions are
full-tree operations such as 'checkout', 'export' and 'import'.

In memory inventories
=====================

In memory inventories are often used in diff and status operations between
trees. We are working to reduce the number of times this occurs with 'full
tree' inventory objects, and instead use more custom tailored data structures
that allow operations on only a small amount of data regardless of the size of
the tree.


Serialization
=============

There are several variants of serialised tree shape in use by bzr. To date
these have been mostly xml based, though plugins have offered non-xml versions.

dirstate
--------

The dirstate file in a working tree includes many different tree shapes - one
for the working tree and one for each parent tree, interleaved to allow
efficient diff and status operations.

xml
---

All the xml serialized forms write to and read from a single byte string, whose
hash is then the inventory validator for the commit object.


Serialization scaling and future designs
========================================

Overall efficiency and scaling is constrained by the bottom level structure
that an inventory is stored as. We have a number of goals we want to achieve:

 1. Allow commit to write less than the full tree's data in to the repository
    in the general case. 
 2. Allow the data that is written to be calculated without examining every
    versioned path in the tree.
 3. Generate the exact same representation for a given inventory regardless of
    the amount of history available.
 4. Allow in memory deltas to be generated directly from the serialised form
    without upcasting to a full in-memory representation or examining every
    path in the tree. Ideally the work performed will be proportional to the
    amount of changes between the trees being compared.
 5. Allow fetch to determine the file texts that need to be pulled to ensure
    that the entire tree can be reconstructed without having to probe every
    path in the tree.
 6. Allow bzr to map paths to file ids without reading the entire serialised
    form. This is something that is used by commands such as merge PATH and
    diff -r X PATH.
 7. Let bzr map file ids to paths without reading the entire serialised form.
    This is used by commands that are presenting output to the user such as
    loggerhead, bzr-search, log FILENAME.
 8. We want a strong validator for inventories which is cheap to generate.
    Specifically we should be able to create the generator for a new commit
    without processing all the data of the basis commit.
 9. Testaments generation is currently size(tree), we would like to create a
    new testament standard which requires less work so that signed commits
    are not significantly slower than regular commits.


We have current performance and memory bugs in log -v, merge, commit, diff -r,
loggerhead and status -r which can be addressed by an inventory system
meeting these goals.

Current situation
-----------------

The xml based implementation we use today layers the inventory as a bytestring
which is stored under a single key; the bytestring is then compressed as a 
delta against the bytestring of its left hand parent by the knit code.

Gap analysis:

 1. Succeeds
 2. Fails - generating a new xml representation needs full tree data.
 3. Succeeds - the inventory layer accesses the bytestring, which is
    deterministic
 4. Fails - we have to reconstruct both inventories as trees and then delta
    the resulting in memory objects.
 5. Partial success - the revision field in the inventory can be scanned for
    in both text-delta and full-bytestring form; other revision values than
    those revisions which are being pulled are by definition absent.
 6. Partially succeeds - with appropriate logic a path<->id map can be generated
    just-in-time, but it is complex and still requires reconstructing the
    entire byte-string.
 7. As for 6.
 8. Fails - we have to hash the entire tree in serialised form to generate
    validators.
 9. Fails.

Long term work
--------------

Some things are likely harder to fix incrementally than others. In particular,
goal 3 (constant canonical form) is arguably only achieved if we remove all
derived data such as the last-modified revision from the inventory itself. That
said, the last-modified appears to be in a higher level than raw serialization.
So in the medium term we will not alter the contents of inventories, only the
way that the current contents are mapped to and from disk.


Layering
--------

We desire clear and clean layers. Each layer should be as simple as we can make
it to aid in debugging and performance tuning. So where we can choose to either
write a complex layer and something simple on top of it, or two layers with
neither being as complex - then we should consider the latter choice better in
the absence of compelling reasons not to.

Some key layers we have today and can look at using or tweaking are:

 * Tree objects - the abstract interface bzrlib code works in
 * VersionedFiles - the optionally delta compressing key->bytes storage
   interface.
 * Inventory - the abstract interface that many tree operations are written in.

These layers are probably sufficient with minor tweaking. We may want to add
additional modules/implementations of one or more layers, but that doesn't
really require new layers to be exposed.

Design elements to achieve the goals in a future inventory implementation
-------------------------------------------------------------------------

 * Split up the logical document into smaller serialised fragements. For
   instance hash buckets or nodes in a tree of some sort. By serialising in 
   smaller units, we can increase the number of smaller units rather than 
   their size as the tree grows; as long as two similar trees have similar
   serialised forms, the amount of different content should be quite high.

 * Use fragment identifiers that are independent of revision id, so that
   serialisation of two related trees generates overlap in the keyspace
   for fragments without requiring explicit delta logic. Content Hash Keys
   (e.g. ('sha1:ABCDEF0123456789...',) are useful here because of the ability
   to assign them without reference to history.)

 * Store the fragments in our existing VersionedFiles store. Adding an index
   for them. Have the serialised form be uncompressed utf8, so that delta logic
   in the VersionedFiles layer can be used. We may need to provide some sort
   of hinting mechanism to get good compression - but the trivially available
   zlib compression of knits-with-no-deltas is probably a good start.

 * Item_keys_introduced_by is innately a history-using function; we can
   reproduce the text-key finding logic by doing a tree diff between any tree
   and an older tree - that will limit the amount of data we need to process
   to something proportional to the difference and the size of each fragment.
   When checking many versions we can track which fragments we have examined
   and only look at new unique ones as each version is examined in turn.

 * Working tree to arbitrary history revision deltas/comparisons can be scaled
   up by doing a two-step (fixed at two!) delta combining - delta(tree, basis)
   and then combine that with delta(basis, arbitrary_revision) using the 
   repositories ability to get a delta cheaply.

 * The key primitives we need seem to be:
   * canonical_form(inventory) -> fragments
   * delta(inventory, inventory) -> inventory_delta
   * apply(inventory_delta, canonical_form) -> fragments

 * Having very many small fragments is likely to cause a high latency
   multiplier unless we are careful.

 * Possible designs to investigate - a hash bucket approach, radix trees,
   B+ trees, directory trees (with splits inside a directory?).


Hash bucket based inventories
=============================

Overview
--------

We store two maps - fileid:inventory_entry and path:fileid, in a stable
hash trie, stored in densly packed fragments. We pack keys into the map
densely up the tree, with a single canonical form for any given tree. This is
more stable than simple fixed size buckets, which prevents corner cases where
the tree size varies right on a bucket size border. (Note that such cases are
not a fatal flaw - the two forms would both be present in the repository, so
only a small amount of data would be written at each transition - but a full
tree reprocess would be needed at each tree operation across the boundary, and
thats undesirable.)

Goal satisfaction
-----------------

 1. Success
 2. Success
 3. Success
 4. Success, though each change will need its parents looked up as well
    so it will be proportional to the changes + the directories above
    the changed path.
 5. Success - looking at the difference against all parents we can determine
    new keys without reference to the repository content will be inserted
    into.
 6. This probably needs a path->id map, allowing a 2-step lookup.
 7. If we allocate buckets by hashing the id, then this is succeed, though,
    as per 4 it will need recursive lookups.
 8. Success
 9. Fail - data beyond that currently included in testaments is included
    in the strong validator.

Issues
------

 1. Tuning the fragment size needs doing.
 1. Testing.
 1. Writing code.
 1. Separate root node, or inline into revision?
 1. Cannot do 'ls' efficiently in the current design.
 1. Cannot detect invalid deltas easily.
 1. What about LCA merge of inventories?

Canonical form
--------------

There are three fragment types for the canonical form. Each fragment is
addressed using a Content Hash Key (CHK) - for instance
"sha1:12345678901234567890".

root_node: (Perhaps this should be inlined into the revision object).
HASH_INVENTORY_SIGNATURE
path_map: CHK to root of path to id map
content_map: CHK to root of id to entry map

map_node: INTERNAL_NODE or LEAF_NODE
INTERNAL_NODE:
INTERNAL_NODE_SIGNATURE
hash_prefix: PREFIX
prefix_width: INT
PREFIX CHK TYPE SIZE
PREFIX CHK TYPE SIZE ...

(Where TYPE is I for internal or L for leaf).

leaf_node:
LEAF_NODE_SIGNATURE
hash_prefix: PREFIX
HASH\x00KEY\x00 VALUE

For path maps, VALUE is::
  fileid

For content maps, VALUE::
  fileid basename kind last-changed kind-specific-details


The path and content maps are populated simply by serialising every inventory
entry and inserting them into both the path map and the content map. The maps
start with just a single leaf node with an empty prefix. 


Apply
-----

Given an inventory delta - a list of (old_path, new_path, InventoryEntry)
items, with a None in new_path indicating a delete operation, and recursive
deletes not being permitted - all entries to be deleted must be explicitly
listed, we can transform a current inventory directly. We can't trivially
detect an invalid delta though.

To perform an application, naively we can just update both maps. For the path
map we would remove all entries where the paths in the delta do not match, then
insert those with a new_path again. For the content map we would just remove
all the fileids in the delta, then insert those with a new_path that is not
None.

Delta
-----

To generate a delta between two inventories, we first generate a list of
altered fileids, and then recursively look up their parents to generate their
old and new file paths.

To generate the list of altered file ids, we do an entry by entry comparison of
the full contents of every leaf node that the two inventories do not have in
common. To do this, we start at the root node, and follow every CHK pointer
that is only in one tree. We can then bring in all the values from the leaf
nodes and do a set difference to get the altered ones, which we would then
parse.


Radix tree based inventories
============================

Overview
--------

We store two maps - fileid:path and path:inventory_entry. The fileid:path map
is a hash trie (as file ids have no useful locality of reference). The
path:inventory_entry map is stored as a regular trie. As for hash tries we
define a single canonical representation for regular tries similar to that
defined above for hash tries.

Goal satisfaction
-----------------

 1. Success
 2. Success
 3. Success
 4. Success
 5. Success - looking at the difference against all parents we can determine
    new keys without reference to the repository content will be inserted
    into.
 6. Success
 7. Success
 8. Success
 9. Fail - data beyond that currently included in testaments is included
    in the strong validator.

Issues
------

 1. Tuning the fragment size needs doing.
 1. Testing.
 1. Writing code.
 1. Separate root node, or inline into revision?
 1. What about LCA merge of inventories?

Canonical form
--------------

There are five fragment types for the canonical form:

The root node, hash trie internal and leaf nodes as previous.

Then we have two more, the internal and leaf node for the radix tree.

radix_node: INTERNAL_NODE or LEAF_NODE

INTERNAL_NODE:
INTERNAL_NODE_SIGNATURE
prefix: PREFIX
suffix CHK TYPE SIZE
suffix CHK TYPE SIZE ...

(Where TYPE is I for internal or L for leaf).

LEAF_NODE:
LEAF_NODE_SIGNATURE
prefix: PREFIX
suffix\x00VALUE

For the content map we use the same value as for hashtrie inventories.


Node splitting and joining in the radix tree are managed in the same fashion as
as for the internal nodes of the hashtries.


Apply
-----

Apply is implemented as for hashtries - we just remove and reinsert the
fileid:paths map entries, and likewise for the path:entry map. We can however
cheaply detect invalid deltas where a delete fails to include its children.

Delta
-----

Delta generation is very similar to that with hash tries, except we get the
path of nodes as part of the lookup process.


Hash Trie details
=================

The canonical form for a hash trie is a tree of internal nodes leading down to
leaf nodes, with no node exceeding some threshold size, and every node
containing as much content as it can, but no leaf node containing less than
its lower size threshold. (In the event that an imbalance in the hash function
causes a tree where an internal node is needed, but any prefix generates a
child with less than the lower threshold, the smallest prefix should be taken).
An internal node holds some number of key prefixes, all with the same bit-width.
A leaf node holds the actual values. As trees do not spring fully-formed, the
canonical form is defined iteratively - by taking every item in a tree and
inserting it into a new tree in order you can determine what canonical form
would look like.  As that is an expensive operation, it should only be done
rarely.

Updates to a tree that is in canonical form can be done preserving canonical
form if we can prove that our rules for insertion are order-independent,
and that our rules for deletion generate the same tree as if we never
inserted those nodes.

Our hash tries are balanced vertically but not horizontally. That is, one leg
of a tree can be arbitrarily deeper than adjacent legs. We require that each
node along a path within the tree be densely packed, with the densest nodes
near the top of the tree, and the least dense at the bottom. Except where the
tree cannot support it, no node is smaller than a minimum_size, and none
larger than maximum_size. The minimum size constraint is only applied when
there are enough entries under a prefix to meet that minimum. The maximum
size constraint is always applied except when a node with a single entry
is larger than the maximum size. Loosely, the maximum size constraint wins
over the minimum size constraint, and if the minimum size contraint is to
be ignored, a deeper prefix can be chosen to pack the containing node more
densely, as long as no additional minimum sizes checks on child nodes are
violated.

Insertion
---------

#. Hash the entry, and insert the entry in the leaf node with a matching
   prefix, creating that node and linking it from the internal node containing
   that prefix if there is no appropriate leaf node.
#. Starting at the highest node altered, for all altered nodes, check if it has
   transitioned across either size boundary - 0 < min_size < max_size. If it
   has not, proceed to update the CHK pointers.
#. If it increased above min_size, check the node above to see if it can be
   more densely packed. To be below the min_size the node's parent must
   have hit the max size constraint and been forced to split even though this
   child did not have enough content to support a min_size node - so the prefix
   chosen in the parent may be shorter than desirable and we may now be able
   to more densely pack the parent by splitting the child nodes more. So if the
   parent node can support a deeper prefix without hitting max_size, and the
   count of under min_size nodes cannot be reduced, the parent should be given
   a deeper prefix.
#. If it increased above max_size, shrink the prefix width used to split out
   new nodes until the node is below max_size (unless the prefix width is
   already 1 - the minimum).
   To shrink the prefix of an internal node, create new internal nodes for each
   new prefix, and populate them with the content of the nodes which were
   formerly linked. (This will normally bubble down due to keeping densely
   packed nodes).
   To shrink the prefix of a leaf node, create an internal node with the same
   prefix, then choose a width for the internal node such that the contents 
   of the leaf all fit into new leaves obeying the min_size and max_size rules.
   The largest prefix possible should be chosen, to obey the
   higher-nodes-are-denser rule. That rule also gives room in leaf nodes for 
   growth without affecting the parent node packing.
#. Update the CHK pointers - serialise every altered node to generate a CHK,
   and update the CHK placeholder in the nodes parent; then reserialise the
   parent. CHK pointer propogation can be done lazily when many updates are
   expected.

Multiple versions of nodes for the same PREFIX and internal prefix width should
compress well for the same tree.