ccan/junkcode/rusty@rustcorp.com.au-ntdb/doc/design.txt

NTDB: Redesigning The Trivial DataBase

Rusty Russell, IBM Corporation

19 June 2012

Abstract

The Trivial DataBase on-disk format is 32 bits; with usage cases
heading towards the 4G limit, that must change. This required
breakage provides an opportunity to revisit TDB's other design
decisions and reassess them.

1 Introduction

The Trivial DataBase was originally written by Andrew Tridgell as
a simple key/data pair storage system with the same API as dbm,
but allowing multiple readers and writers while being small
enough (< 1000 lines of C) to include in SAMBA. The simple design
created in 1999 has proven surprisingly robust and performant,
used in Samba versions 3 and 4 as well as numerous other
projects. Its useful life was greatly increased by the
(backwards-compatible!) addition of transaction support in 2005.

The wider variety and greater demands of TDB-using code has lead
to some organic growth of the API, as well as some compromises on
the implementation. None of these, by themselves, are seen as
show-stoppers, but the cumulative effect is to a loss of elegance
over the initial, simple TDB implementation. Here is a table of
the approximate number of lines of implementation code and number
of API functions at the end of each year:


+-----------+----------------+--------------------------------+
| Year End  | API Functions  | Lines of C Code Implementation |
+-----------+----------------+--------------------------------+
+-----------+----------------+--------------------------------+
|   1999    |      13        |              1195              |
+-----------+----------------+--------------------------------+
|   2000    |      24        |              1725              |
+-----------+----------------+--------------------------------+
|   2001    |      32        |              2228              |
+-----------+----------------+--------------------------------+
|   2002    |      35        |              2481              |
+-----------+----------------+--------------------------------+
|   2003    |      35        |              2552              |
+-----------+----------------+--------------------------------+
|   2004    |      40        |              2584              |
+-----------+----------------+--------------------------------+
|   2005    |      38        |              2647              |
+-----------+----------------+--------------------------------+
|   2006    |      52        |              3754              |
+-----------+----------------+--------------------------------+
|   2007    |      66        |              4398              |
+-----------+----------------+--------------------------------+
|   2008    |      71        |              4768              |
+-----------+----------------+--------------------------------+
|   2009    |      73        |              5715              |
+-----------+----------------+--------------------------------+


This review is an attempt to catalog and address all the known
issues with TDB and create solutions which address the problems
without significantly increasing complexity; all involved are far
too aware of the dangers of second system syndrome in rewriting a
successful project like this.

Note: the final decision was to make ntdb a separate library,
with a separarate 'ntdb' namespace so both can potentially be
linked together. This document still refers to “tdb” everywhere,
for simplicity.

2 API Issues

2.1 tdb_open_ex Is Not Expandable

The tdb_open() call was expanded to tdb_open_ex(), which added an
optional hashing function and an optional logging function
argument. Additional arguments to open would require the
introduction of a tdb_open_ex2 call etc.

2.1.1 Proposed Solution<attributes>

tdb_open() will take a linked-list of attributes:

enum tdb_attribute {

    TDB_ATTRIBUTE_LOG = 0,

    TDB_ATTRIBUTE_HASH = 1

};

struct tdb_attribute_base {

    enum tdb_attribute attr;

    union tdb_attribute *next;

};

struct tdb_attribute_log {

    struct tdb_attribute_base base; /* .attr = TDB_ATTRIBUTE_LOG
*/

    tdb_log_func log_fn;

    void *log_private;

};

struct tdb_attribute_hash {

    struct tdb_attribute_base base; /* .attr = TDB_ATTRIBUTE_HASH
*/

    tdb_hash_func hash_fn;

    void *hash_private;

};

union tdb_attribute {

    struct tdb_attribute_base base;

    struct tdb_attribute_log log;

    struct tdb_attribute_hash hash;

};

This allows future attributes to be added, even if this expands
the size of the union.

2.1.2 Status

Complete.

2.2 tdb_traverse Makes Impossible Guarantees

tdb_traverse (and tdb_firstkey/tdb_nextkey) predate transactions,
and it was thought that it was important to guarantee that all
records which exist at the start and end of the traversal would
be included, and no record would be included twice.

This adds complexity (see[Reliable-Traversal-Adds]) and does not
work anyway for records which are altered (in particular, those
which are expanded may be effectively deleted and re-added behind
the traversal).

2.2.1 <traverse-Proposed-Solution>Proposed Solution

Abandon the guarantee. You will see every record if no changes
occur during your traversal, otherwise you will see some subset.
You can prevent changes by using a transaction or the locking
API.

2.2.2 Status

Complete. Delete-during-traverse will still delete every record,
too (assuming no other changes).

2.3 Nesting of Transactions Is Fraught

TDB has alternated between allowing nested transactions and not
allowing them. Various paths in the Samba codebase assume that
transactions will nest, and in a sense they can: the operation is
only committed to disk when the outer transaction is committed.
There are two problems, however:

1. Canceling the inner transaction will cause the outer
  transaction commit to fail, and will not undo any operations
  since the inner transaction began. This problem is soluble with
  some additional internal code.

2. An inner transaction commit can be cancelled by the outer
  transaction. This is desirable in the way which Samba's
  database initialization code uses transactions, but could be a
  surprise to any users expecting a successful transaction commit
  to expose changes to others.

The current solution is to specify the behavior at tdb_open(),
with the default currently that nested transactions are allowed.
This flag can also be changed at runtime.

2.3.1 Proposed Solution

Given the usage patterns, it seems that the“least-surprise”
behavior of disallowing nested transactions should become the
default. Additionally, it seems the outer transaction is the only
code which knows whether inner transactions should be allowed, so
a flag to indicate this could be added to tdb_transaction_start.
However, this behavior can be simulated with a wrapper which uses
tdb_add_flags() and tdb_remove_flags(), so the API should not be
expanded for this relatively-obscure case.

2.3.2 Status

Complete; the nesting flag has been removed.

2.4 Incorrect Hash Function is Not Detected

tdb_open_ex() allows the calling code to specify a different hash
function to use, but does not check that all other processes
accessing this tdb are using the same hash function. The result
is that records are missing from tdb_fetch().

2.4.1 Proposed Solution

The header should contain an example hash result (eg. the hash of
0xdeadbeef), and tdb_open_ex() should check that the given hash
function produces the same answer, or fail the tdb_open call.

2.4.2 Status

Complete.

2.5 tdb_set_max_dead/TDB_VOLATILE Expose Implementation

In response to scalability issues with the free list ([TDB-Freelist-Is]
) two API workarounds have been incorporated in TDB:
tdb_set_max_dead() and the TDB_VOLATILE flag to tdb_open. The
latter actually calls the former with an argument of“5”.

This code allows deleted records to accumulate without putting
them in the free list. On delete we iterate through each chain
and free them in a batch if there are more than max_dead entries.
These are never otherwise recycled except as a side-effect of a
tdb_repack.

2.5.1 Proposed Solution

With the scalability problems of the freelist solved, this API
can be removed. The TDB_VOLATILE flag may still be useful as a
hint that store and delete of records will be at least as common
as fetch in order to allow some internal tuning, but initially
will become a no-op.

2.5.2 Status

Complete. Unknown flags cause tdb_open() to fail as well, so they
can be detected at runtime.

2.6 <TDB-Files-Cannot>TDB Files Cannot Be Opened Multiple Times
  In The Same Process

No process can open the same TDB twice; we check and disallow it.
This is an unfortunate side-effect of fcntl locks, which operate
on a per-file rather than per-file-descriptor basis, and do not
nest. Thus, closing any file descriptor on a file clears all the
locks obtained by this process, even if they were placed using a
different file descriptor!

Note that even if this were solved, deadlock could occur if
operations were nested: this is a more manageable programming
error in most cases.

2.6.1 Proposed Solution

We could lobby POSIX to fix the perverse rules, or at least lobby
Linux to violate them so that the most common implementation does
not have this restriction. This would be a generally good idea
for other fcntl lock users.

Samba uses a wrapper which hands out the same tdb_context to
multiple callers if this happens, and does simple reference
counting. We should do this inside the tdb library, which already
emulates lock nesting internally; it would need to recognize when
deadlock occurs within a single process. This would create a new
failure mode for tdb operations (while we currently handle
locking failures, they are impossible in normal use and a process
encountering them can do little but give up).

I do not see benefit in an additional tdb_open flag to indicate
whether re-opening is allowed, as though there may be some
benefit to adding a call to detect when a tdb_context is shared,
to allow other to create such an API.

2.6.2 Status

Complete.

2.7 TDB API Is Not POSIX Thread-safe

The TDB API uses an error code which can be queried after an
operation to determine what went wrong. This programming model
does not work with threads, unless specific additional guarantees
are given by the implementation. In addition, even
otherwise-independent threads cannot open the same TDB (as in[TDB-Files-Cannot]
).

2.7.1 Proposed Solution

Reachitecting the API to include a tdb_errcode pointer would be a
great deal of churn, but fortunately most functions return 0 on
success and -1 on error: we can change these to return 0 on
success and a negative error code on error, and the API remains
similar to previous. The tdb_fetch, tdb_firstkey and tdb_nextkey
functions need to take a TDB_DATA pointer and return an error
code. It is also simpler to have tdb_nextkey replace its key
argument in place, freeing up any old .dptr.

Internal locking is required to make sure that fcntl locks do not
overlap between threads, and also that the global list of tdbs is
maintained.

The aim is that building tdb with -DTDB_PTHREAD will result in a
pthread-safe version of the library, and otherwise no overhead
will exist. Alternatively, a hooking mechanism similar to that
proposed for[Proposed-Solution-locking-hook] could be used to
enable pthread locking at runtime.

2.7.2 Status

Incomplete; API has been changed but thread safety has not been
implemented.

2.8 *_nonblock Functions And *_mark Functions Expose
  Implementation

CTDB[footnote:
Clustered TDB, see http://ctdb.samba.org
] wishes to operate on TDB in a non-blocking manner. This is
currently done as follows:

1. Call the _nonblock variant of an API function (eg.
  tdb_lockall_nonblock). If this fails:

2. Fork a child process, and wait for it to call the normal
  variant (eg. tdb_lockall).

3. If the child succeeds, call the _mark variant to indicate we
  already have the locks (eg. tdb_lockall_mark).

4. Upon completion, tell the child to release the locks (eg.
  tdb_unlockall).

5. Indicate to tdb that it should consider the locks removed (eg.
  tdb_unlockall_mark).

There are several issues with this approach. Firstly, adding two
new variants of each function clutters the API for an obscure
use, and so not all functions have three variants. Secondly, it
assumes that all paths of the functions ask for the same locks,
otherwise the parent process will have to get a lock which the
child doesn't have under some circumstances. I don't believe this
is currently the case, but it constrains the implementation.

2.8.1 <Proposed-Solution-locking-hook>Proposed Solution

Implement a hook for locking methods, so that the caller can
control the calls to create and remove fcntl locks. In this
scenario, ctdbd would operate as follows:

1. Call the normal API function, eg tdb_lockall().

2. When the lock callback comes in, check if the child has the
  lock. Initially, this is always false. If so, return 0.
  Otherwise, try to obtain it in non-blocking mode. If that
  fails, return EWOULDBLOCK.

3. Release locks in the unlock callback as normal.

4. If tdb_lockall() fails, see if we recorded a lock failure; if
  so, call the child to repeat the operation.

5. The child records what locks it obtains, and returns that
  information to the parent.

6. When the child has succeeded, goto 1.

This is flexible enough to handle any potential locking scenario,
even when lock requirements change. It can be optimized so that
the parent does not release locks, just tells the child which
locks it doesn't need to obtain.

It also keeps the complexity out of the API, and in ctdbd where
it is needed.

2.8.2 Status

Complete.

2.9 tdb_chainlock Functions Expose Implementation

tdb_chainlock locks some number of records, including the record
indicated by the given key. This gave atomicity guarantees;
no-one can start a transaction, alter, read or delete that key
while the lock is held.

It also makes the same guarantee for any other key in the chain,
which is an internal implementation detail and potentially a
cause for deadlock.

2.9.1 Proposed Solution

None. It would be nice to have an explicit single entry lock
which effected no other keys. Unfortunately, this won't work for
an entry which doesn't exist. Thus while chainlock may be
implemented more efficiently for the existing case, it will still
have overlap issues with the non-existing case. So it is best to
keep the current (lack of) guarantee about which records will be
effected to avoid constraining our implementation.

2.10 Signal Handling is Not Race-Free

The tdb_setalarm_sigptr() call allows the caller's signal handler
to indicate that the tdb locking code should return with a
failure, rather than trying again when a signal is received (and
errno == EAGAIN). This is usually used to implement timeouts.

Unfortunately, this does not work in the case where the signal is
received before the tdb code enters the fcntl() call to place the
lock: the code will sleep within the fcntl() code, unaware that
the signal wants it to exit. In the case of long timeouts, this
does not happen in practice.

2.10.1 Proposed Solution

The locking hooks proposed in[Proposed-Solution-locking-hook]
would allow the user to decide on whether to fail the lock
acquisition on a signal. This allows the caller to choose their
own compromise: they could narrow the race by checking
immediately before the fcntl call.[footnote:
It may be possible to make this race-free in some implementations
by having the signal handler alter the struct flock to make it
invalid. This will cause the fcntl() lock call to fail with
EINVAL if the signal occurs before the kernel is entered,
otherwise EAGAIN.
]

2.10.2 Status

Complete.

2.11 The API Uses Gratuitous Typedefs, Capitals

typedefs are useful for providing source compatibility when types
can differ across implementations, or arguably in the case of
function pointer definitions which are hard for humans to parse.
Otherwise it is simply obfuscation and pollutes the namespace.

Capitalization is usually reserved for compile-time constants and
macros.

  TDB_CONTEXT There is no reason to use this over 'struct
  tdb_context'; the definition isn't visible to the API user
  anyway.

  TDB_DATA There is no reason to use this over struct TDB_DATA;
  the struct needs to be understood by the API user.

  struct TDB_DATA This would normally be called 'struct
  tdb_data'.

  enum TDB_ERROR Similarly, this would normally be enum
  tdb_error.

2.11.1 Proposed Solution

None. Introducing lower case variants would please pedants like
myself, but if it were done the existing ones should be kept.
There is little point forcing a purely cosmetic change upon tdb
users.

2.12 <tdb_log_func-Doesnt-Take>tdb_log_func Doesn't Take The
  Private Pointer

For API compatibility reasons, the logging function needs to call
tdb_get_logging_private() to retrieve the pointer registered by
the tdb_open_ex for logging.

2.12.1 Proposed Solution

It should simply take an extra argument, since we are prepared to
break the API/ABI.

2.12.2 Status

Complete.

2.13 Various Callback Functions Are Not Typesafe

The callback functions in tdb_set_logging_function (after[tdb_log_func-Doesnt-Take]
 is resolved), tdb_parse_record, tdb_traverse, tdb_traverse_read
and tdb_check all take void * and must internally convert it to
the argument type they were expecting.

If this type changes, the compiler will not produce warnings on
the callers, since it only sees void *.

2.13.1 Proposed Solution

With careful use of macros, we can create callback functions
which give a warning when used on gcc and the types of the
callback and its private argument differ. Unsupported compilers
will not give a warning, which is no worse than now. In addition,
the callbacks become clearer, as they need not use void * for
their parameter.

See CCAN's typesafe_cb module at
http://ccan.ozlabs.org/info/typesafe_cb.html

2.13.2 Status

Complete.

2.14 TDB_CLEAR_IF_FIRST Must Be Specified On All Opens,
  tdb_reopen_all Problematic

The TDB_CLEAR_IF_FIRST flag to tdb_open indicates that the TDB
file should be cleared if the caller discovers it is the only
process with the TDB open. However, if any caller does not
specify TDB_CLEAR_IF_FIRST it will not be detected, so will have
the TDB erased underneath them (usually resulting in a crash).

There is a similar issue on fork(); if the parent exits (or
otherwise closes the tdb) before the child calls tdb_reopen_all()
to establish the lock used to indicate the TDB is opened by
someone, a TDB_CLEAR_IF_FIRST opener at that moment will believe
it alone has opened the TDB and will erase it.

2.14.1 Proposed Solution

Remove TDB_CLEAR_IF_FIRST. Other workarounds are possible, but
see[TDB_CLEAR_IF_FIRST-Imposes-Performance].

2.14.2 Status

Complete. An open hook is provided to replicate this
functionality if required.

2.15 Extending The Header Is Difficult

We have reserved (zeroed) words in the TDB header, which can be
used for future features. If the future features are compulsory,
the version number must be updated to prevent old code from
accessing the database. But if the future feature is optional, we
have no way of telling if older code is accessing the database or
not.

2.15.1 Proposed Solution

The header should contain a“format variant” value (64-bit). This
is divided into two 32-bit parts:

1. The lower part reflects the format variant understood by code
  accessing the database.

2. The upper part reflects the format variant you must understand
  to write to the database (otherwise you can only open for
  reading).

The latter field can only be written at creation time, the former
should be written under the OPEN_LOCK when opening the database
for writing, if the variant of the code is lower than the current
lowest variant.

This should allow backwards-compatible features to be added, and
detection if older code (which doesn't understand the feature)
writes to the database.

2.15.2 Status

Complete.

2.16 Record Headers Are Not Expandible

If we later want to add (say) checksums on keys and data, it
would require another format change, which we'd like to avoid.

2.16.1 Proposed Solution

We often have extra padding at the tail of a record. If we ensure
that the first byte (if any) of this padding is zero, we will
have a way for future changes to detect code which doesn't
understand a new format: the new code would write (say) a 1 at
the tail, and thus if there is no tail or the first byte is 0, we
would know the extension is not present on that record.

2.16.2 Status

Complete.

2.17 TDB Does Not Use Talloc

Many users of TDB (particularly Samba) use the talloc allocator,
and thus have to wrap TDB in a talloc context to use it
conveniently.

2.17.1 Proposed Solution

The allocation within TDB is not complicated enough to justify
the use of talloc, and I am reluctant to force another
(excellent) library on TDB users. Nonetheless a compromise is
possible. An attribute (see[attributes]) can be added later to
tdb_open() to provide an alternate allocation mechanism,
specifically for talloc but usable by any other allocator (which
would ignore the“context” argument).

This would form a talloc heirarchy as expected, but the caller
would still have to attach a destructor to the tdb context
returned from tdb_open to close it. All TDB_DATA fields would be
children of the tdb_context, and the caller would still have to
manage them (using talloc_free() or talloc_steal()).

2.17.2 Status

Complete, using the NTDB_ATTRIBUTE_ALLOCATOR attribute.

3 Performance And Scalability Issues

3.1 <TDB_CLEAR_IF_FIRST-Imposes-Performance>TDB_CLEAR_IF_FIRST
  Imposes Performance Penalty

When TDB_CLEAR_IF_FIRST is specified, a 1-byte read lock is
placed at offset 4 (aka. the ACTIVE_LOCK). While these locks
never conflict in normal tdb usage, they do add substantial
overhead for most fcntl lock implementations when the kernel
scans to detect if a lock conflict exists. This is often a single
linked list, making the time to acquire and release a fcntl lock
O(N) where N is the number of processes with the TDB open, not
the number actually doing work.

In a Samba server it is common to have huge numbers of clients
sitting idle, and thus they have weaned themselves off the
TDB_CLEAR_IF_FIRST flag.[footnote:
There is a flag to tdb_reopen_all() which is used for this
optimization: if the parent process will outlive the child, the
child does not need the ACTIVE_LOCK. This is a workaround for
this very performance issue.
]

3.1.1 Proposed Solution

Remove the flag. It was a neat idea, but even trivial servers
tend to know when they are initializing for the first time and
can simply unlink the old tdb at that point.

3.1.2 Status

Complete.

3.2 TDB Files Have a 4G Limit

This seems to be becoming an issue (so much for“trivial”!),
particularly for ldb.

3.2.1 Proposed Solution

A new, incompatible TDB format which uses 64 bit offsets
internally rather than 32 bit as now. For simplicity of endian
conversion (which TDB does on the fly if required), all values
will be 64 bit on disk. In practice, some upper bits may be used
for other purposes, but at least 56 bits will be available for
file offsets.

tdb_open() will automatically detect the old version, and even
create them if TDB_VERSION6 is specified to tdb_open.

32 bit processes will still be able to access TDBs larger than 4G
(assuming that their off_t allows them to seek to 64 bits), they
will gracefully fall back as they fail to mmap. This can happen
already with large TDBs.

Old versions of tdb will fail to open the new TDB files (since 28
August 2009, commit 398d0c29290: prior to that any unrecognized
file format would be erased and initialized as a fresh tdb!)

3.2.2 Status

Complete.

3.3 TDB Records Have a 4G Limit

This has not been a reported problem, and the API uses size_t
which can be 64 bit on 64 bit platforms. However, other limits
may have made such an issue moot.

3.3.1 Proposed Solution

Record sizes will be 64 bit, with an error returned on 32 bit
platforms which try to access such records (the current
implementation would return TDB_ERR_OOM in a similar case). It
seems unlikely that 32 bit keys will be a limitation, so the
implementation may not support this (see[sub:Records-Incur-A]).

3.3.2 Status

Complete.

3.4 Hash Size Is Determined At TDB Creation Time

TDB contains a number of hash chains in the header; the number is
specified at creation time, and defaults to 131. This is such a
bottleneck on large databases (as each hash chain gets quite
long), that LDB uses 10,000 for this hash. In general it is
impossible to know what the 'right' answer is at database
creation time.

3.4.1 <sub:Hash-Size-Solution>Proposed Solution

After comprehensive performance testing on various scalable hash
variants[footnote:
http://rusty.ozlabs.org/?p=89 and http://rusty.ozlabs.org/?p=94
This was annoying because I was previously convinced that an
expanding tree of hashes would be very close to optimal.
], it became clear that it is hard to beat a straight linear hash
table which doubles in size when it reaches saturation.
Unfortunately, altering the hash table introduces serious locking
complications: the entire hash table needs to be locked to
enlarge the hash table, and others might be holding locks.
Particularly insidious are insertions done under tdb_chainlock.

Thus an expanding layered hash will be used: an array of hash
groups, with each hash group exploding into pointers to lower
hash groups once it fills, turning into a hash tree. This has
implications for locking: we must lock the entire group in case
we need to expand it, yet we don't know how deep the tree is at
that point.

Note that bits from the hash table entries should be stolen to
hold more hash bits to reduce the penalty of collisions. We can
use the otherwise-unused lower 3 bits. If we limit the size of
the database to 64 exabytes, we can use the top 8 bits of the
hash entry as well. These 11 bits would reduce false positives
down to 1 in 2000 which is more than we need: we can use one of
the bits to indicate that the extra hash bits are valid. This
means we can choose not to re-hash all entries when we expand a
hash group; simply use the next bits we need and mark them
invalid.

3.4.2 Status

Ignore. Scaling the hash automatically proved inefficient at
small hash sizes; we default to a 8192-element hash (changable
via NTDB_ATTRIBUTE_HASHSIZE), and when buckets clash we expand to
an array of hash entries. This scales slightly better than the
tdb chain (due to the 8 top bits containing extra hash).

3.5 <TDB-Freelist-Is>TDB Freelist Is Highly Contended

TDB uses a single linked list for the free list. Allocation
occurs as follows, using heuristics which have evolved over time:

1. Get the free list lock for this whole operation.

2. Multiply length by 1.25, so we always over-allocate by 25%.

3. Set the slack multiplier to 1.

4. Examine the current freelist entry: if it is > length but <
  the current best case, remember it as the best case.

5. Multiply the slack multiplier by 1.05.

6. If our best fit so far is less than length * slack multiplier,
  return it. The slack will be turned into a new free record if
  it's large enough.

7. Otherwise, go onto the next freelist entry.

Deleting a record occurs as follows:

1. Lock the hash chain for this whole operation.

2. Walk the chain to find the record, keeping the prev pointer
  offset.

3. If max_dead is non-zero:

  (a) Walk the hash chain again and count the dead records.

  (b) If it's more than max_dead, bulk free all the dead ones
    (similar to steps 4 and below, but the lock is only obtained
    once).

  (c) Simply mark this record as dead and return.

4. Get the free list lock for the remainder of this operation.

5. <right-merging>Examine the following block to see if it is
  free; if so, enlarge the current block and remove that block
  from the free list. This was disabled, as removal from the free
  list was O(entries-in-free-list).

6. Examine the preceeding block to see if it is free: for this
  reason, each block has a 32-bit tailer which indicates its
  length. If it is free, expand it to cover our new block and
  return.

7. Otherwise, prepend ourselves to the free list.

Disabling right-merging (step[right-merging]) causes
fragmentation; the other heuristics proved insufficient to
address this, so the final answer to this was that when we expand
the TDB file inside a transaction commit, we repack the entire
tdb.

The single list lock limits our allocation rate; due to the other
issues this is not currently seen as a bottleneck.

3.5.1 Proposed Solution

The first step is to remove all the current heuristics, as they
obviously interact, then examine them once the lock contention is
addressed.

The free list must be split to reduce contention. Assuming
perfect free merging, we can at most have 1 free list entry for
each entry. This implies that the number of free lists is related
to the size of the hash table, but as it is rare to walk a large
number of free list entries we can use far fewer, say 1/32 of the
number of hash buckets.

It seems tempting to try to reuse the hash implementation which
we use for records here, but we have two ways of searching for
free entries: for allocation we search by size (and possibly
zone) which produces too many clashes for our hash table to
handle well, and for coalescing we search by address. Thus an
array of doubly-linked free lists seems preferable.

There are various benefits in using per-size free lists (see[sub:TDB-Becomes-Fragmented]
) but it's not clear this would reduce contention in the common
case where all processes are allocating/freeing the same size.
Thus we almost certainly need to divide in other ways: the most
obvious is to divide the file into zones, and using a free list
(or table of free lists) for each. This approximates address
ordering.

Unfortunately it is difficult to know what heuristics should be
used to determine zone sizes, and our transaction code relies on
being able to create a“recovery area” by simply appending to the
file (difficult if it would need to create a new zone header).
Thus we use a linked-list of free tables; currently we only ever
create one, but if there is more than one we choose one at random
to use. In future we may use heuristics to add new free tables on
contention. We only expand the file when all free tables are
exhausted.

The basic algorithm is as follows. Freeing is simple:

1. Identify the correct free list.

2. Lock the corresponding list.

3. Re-check the list (we didn't have a lock, sizes could have
  changed): relock if necessary.

4. Place the freed entry in the list.

Allocation is a little more complicated, as we perform delayed
coalescing at this point:

1. Pick a free table; usually the previous one.

2. Lock the corresponding list.

3. If the top entry is -large enough, remove it from the list and
  return it.

4. Otherwise, coalesce entries in the list.If there was no entry
  large enough, unlock the list and try the next largest list

5. If no list has an entry which meets our needs, try the next
  free table.

6. If no zone satisfies, expand the file.

This optimizes rapid insert/delete of free list entries by not
coalescing them all the time.. First-fit address ordering
ordering seems to be fairly good for keeping fragmentation low
(see[sub:TDB-Becomes-Fragmented]). Note that address ordering
does not need a tailer to coalesce, though if we needed one we
could have one cheaply: see[sub:Records-Incur-A].

Each free entry has the free table number in the header: less
than 255. It also contains a doubly-linked list for easy
deletion.

3.6 <sub:TDB-Becomes-Fragmented>TDB Becomes Fragmented

Much of this is a result of allocation strategy[footnote:
The Memory Fragmentation Problem: Solved? Johnstone & Wilson 1995
ftp://ftp.cs.utexas.edu/pub/garbage/malloc/ismm98.ps
] and deliberate hobbling of coalescing; internal fragmentation
(aka overallocation) is deliberately set at 25%, and external
fragmentation is only cured by the decision to repack the entire
db when a transaction commit needs to enlarge the file.

3.6.1 Proposed Solution

The 25% overhead on allocation works in practice for ldb because
indexes tend to expand by one record at a time. This internal
fragmentation can be resolved by having an“expanded” bit in the
header to note entries that have previously expanded, and
allocating more space for them.

There are is a spectrum of possible solutions for external
fragmentation: one is to use a fragmentation-avoiding allocation
strategy such as best-fit address-order allocator. The other end
of the spectrum would be to use a bump allocator (very fast and
simple) and simply repack the file when we reach the end.

There are three problems with efficient fragmentation-avoiding
allocators: they are non-trivial, they tend to use a single free
list for each size, and there's no evidence that tdb allocation
patterns will match those recorded for general allocators (though
it seems likely).

Thus we don't spend too much effort on external fragmentation; we
will be no worse than the current code if we need to repack on
occasion. More effort is spent on reducing freelist contention,
and reducing overhead.

3.7 <sub:Records-Incur-A>Records Incur A 28-Byte Overhead

Each TDB record has a header as follows:

struct tdb_record {

        tdb_off_t next; /* offset of the next record in the list
*/

        tdb_len_t rec_len; /* total byte length of record */

        tdb_len_t key_len; /* byte length of key */

        tdb_len_t data_len; /* byte length of data */

        uint32_t full_hash; /* the full 32 bit hash of the key */

        uint32_t magic;   /* try to catch errors */

        /* the following union is implied:

                union {

                        char record[rec_len];

                        struct {

                                char key[key_len];

                                char data[data_len];

                        }

                        uint32_t totalsize; (tailer)

                }

        */

};

Naively, this would double to a 56-byte overhead on a 64 bit
implementation.

3.7.1 Proposed Solution

We can use various techniques to reduce this for an allocated
block:

1. The 'next' pointer is not required, as we are using a flat
  hash table.

2. 'rec_len' can instead be expressed as an addition to key_len
  and data_len (it accounts for wasted or overallocated length in
  the record). Since the record length is always a multiple of 8,
  we can conveniently fit it in 32 bits (representing up to 35
  bits).

3. 'key_len' and 'data_len' can be reduced. I'm unwilling to
  restrict 'data_len' to 32 bits, but instead we can combine the
  two into one 64-bit field and using a 5 bit value which
  indicates at what bit to divide the two. Keys are unlikely to
  scale as fast as data, so I'm assuming a maximum key size of 32
  bits.

4. 'full_hash' is used to avoid a memcmp on the“miss” case, but
  this is diminishing returns after a handful of bits (at 10
  bits, it reduces 99.9% of false memcmp). As an aside, as the
  lower bits are already incorporated in the hash table
  resolution, the upper bits should be used here. Note that it's
  not clear that these bits will be a win, given the extra bits
  in the hash table itself (see[sub:Hash-Size-Solution]).

5. 'magic' does not need to be enlarged: it currently reflects
  one of 5 values (used, free, dead, recovery, and
  unused_recovery). It is useful for quick sanity checking
  however, and should not be eliminated.

6. 'tailer' is only used to coalesce free blocks (so a block to
  the right can find the header to check if this block is free).
  This can be replaced by a single 'free' bit in the header of
  the following block (and the tailer only exists in free
  blocks).[footnote:
This technique from Thomas Standish. Data Structure Techniques.
Addison-Wesley, Reading, Massachusetts, 1980.
] The current proposed coalescing algorithm doesn't need this,
  however.

This produces a 16 byte used header like this:

struct tdb_used_record {

        uint32_t used_magic : 16,



                 key_data_divide: 5,

                 top_hash: 11;

        uint32_t extra_octets;

        uint64_t key_and_data_len;

};

And a free record like this:

struct tdb_free_record {

        uint64_t free_magic: 8,

                   prev : 56;



        uint64_t free_table: 8,

                 total_length : 56

        uint64_t next;;

};

Note that by limiting valid offsets to 56 bits, we can pack
everything we need into 3 64-byte words, meaning our minimum
record size is 8 bytes.

3.7.2 Status

Complete.

3.8 Transaction Commit Requires 4 fdatasync

The current transaction algorithm is:

1. write_recovery_data();

2. sync();

3. write_recovery_header();

4. sync();

5. overwrite_with_new_data();

6. sync();

7. remove_recovery_header();

8. sync();

On current ext3, each sync flushes all data to disk, so the next
3 syncs are relatively expensive. But this could become a
performance bottleneck on other filesystems such as ext4.

3.8.1 Proposed Solution

Neil Brown points out that this is overzealous, and only one sync
is needed:

1. Bundle the recovery data, a transaction counter and a strong
  checksum of the new data.

2. Strong checksum that whole bundle.

3. Store the bundle in the database.

4. Overwrite the oldest of the two recovery pointers in the
  header (identified using the transaction counter) with the
  offset of this bundle.

5. sync.

6. Write the new data to the file.

Checking for recovery means identifying the latest bundle with a
valid checksum and using the new data checksum to ensure that it
has been applied. This is more expensive than the current check,
but need only be done at open. For running databases, a separate
header field can be used to indicate a transaction in progress;
we need only check for recovery if this is set.

3.8.2 Status

Deferred.

3.9 <sub:TDB-Does-Not>TDB Does Not Have Snapshot Support

3.9.1 Proposed Solution

None. At some point you say“use a real database” (but see[replay-attribute]
).

But as a thought experiment, if we implemented transactions to
only overwrite free entries (this is tricky: there must not be a
header in each entry which indicates whether it is free, but use
of presence in metadata elsewhere), and a pointer to the hash
table, we could create an entirely new commit without destroying
existing data. Then it would be easy to implement snapshots in a
similar way.

This would not allow arbitrary changes to the database, such as
tdb_repack does, and would require more space (since we have to
preserve the current and future entries at once). If we used hash
trees rather than one big hash table, we might only have to
rewrite some sections of the hash, too.

We could then implement snapshots using a similar method, using
multiple different hash tables/free tables.

3.9.2 Status

Deferred.

3.10 Transactions Cannot Operate in Parallel

This would be useless for ldb, as it hits the index records with
just about every update. It would add significant complexity in
resolving clashes, and cause the all transaction callers to write
their code to loop in the case where the transactions spuriously
failed.

3.10.1 Proposed Solution

None (but see[replay-attribute]). We could solve a small part of
the problem by providing read-only transactions. These would
allow one write transaction to begin, but it could not commit
until all r/o transactions are done. This would require a new
RO_TRANSACTION_LOCK, which would be upgraded on commit.

3.10.2 Status

Deferred.

3.11 Default Hash Function Is Suboptimal

The Knuth-inspired multiplicative hash used by tdb is fairly slow
(especially if we expand it to 64 bits), and works best when the
hash bucket size is a prime number (which also means a slow
modulus). In addition, it is highly predictable which could
potentially lead to a Denial of Service attack in some TDB uses.

3.11.1 Proposed Solution

The Jenkins lookup3 hash[footnote:
http://burtleburtle.net/bob/c/lookup3.c
] is a fast and superbly-mixing hash. It's used by the Linux
kernel and almost everything else. This has the particular
properties that it takes an initial seed, and produces two 32 bit
hash numbers, which we can combine into a 64-bit hash.

The seed should be created at tdb-creation time from some random
source, and placed in the header. This is far from foolproof, but
adds a little bit of protection against hash bombing.

3.11.2 Status

Complete.

3.12 <Reliable-Traversal-Adds>Reliable Traversal Adds Complexity

We lock a record during traversal iteration, and try to grab that
lock in the delete code. If that grab on delete fails, we simply
mark it deleted and continue onwards; traversal checks for this
condition and does the delete when it moves off the record.

If traversal terminates, the dead record may be left
indefinitely.

3.12.1 Proposed Solution

Remove reliability guarantees; see[traverse-Proposed-Solution].

3.12.2 Status

Complete.

3.13 Fcntl Locking Adds Overhead

Placing a fcntl lock means a system call, as does removing one.
This is actually one reason why transactions can be faster
(everything is locked once at transaction start). In the
uncontended case, this overhead can theoretically be eliminated.

3.13.1 Proposed Solution

None.

We tried this before with spinlock support, in the early days of
TDB, and it didn't make much difference except in manufactured
benchmarks.

We could use spinlocks (with futex kernel support under Linux),
but it means that we lose automatic cleanup when a process dies
with a lock. There is a method of auto-cleanup under Linux, but
it's not supported by other operating systems. We could
reintroduce a clear-if-first-style lock and sweep for dead
futexes on open, but that wouldn't help the normal case of one
concurrent opener dying. Increasingly elaborate repair schemes
could be considered, but they require an ABI change (everyone
must use them) anyway, so there's no need to do this at the same
time as everything else.

3.14 Some Transactions Don't Require Durability

Volker points out that gencache uses a CLEAR_IF_FIRST tdb for
normal (fast) usage, and occasionally empties the results into a
transactional TDB. This kind of usage prioritizes performance
over durability: as long as we are consistent, data can be lost.

This would be more neatly implemented inside tdb: a“soft”
transaction commit (ie. syncless) which meant that data may be
reverted on a crash.

3.14.1 Proposed Solution

None.

Unfortunately any transaction scheme which overwrites old data
requires a sync before that overwrite to avoid the possibility of
corruption.

It seems possible to use a scheme similar to that described in[sub:TDB-Does-Not]
,where transactions are committed without overwriting existing
data, and an array of top-level pointers were available in the
header. If the transaction is“soft” then we would not need a sync
at all: existing processes would pick up the new hash table and
free list and work with that.

At some later point, a sync would allow recovery of the old data
into the free lists (perhaps when the array of top-level pointers
filled). On crash, tdb_open() would examine the array of top
levels, and apply the transactions until it encountered an
invalid checksum.

3.15 Tracing Is Fragile, Replay Is External

The current TDB has compile-time-enabled tracing code, but it
often breaks as it is not enabled by default. In a similar way,
the ctdb code has an external wrapper which does replay tracing
so it can coordinate cluster-wide transactions.

3.15.1 Proposed Solution<replay-attribute>

Tridge points out that an attribute can be later added to
tdb_open (see[attributes]) to provide replay/trace hooks, which
could become the basis for this and future parallel transactions
and snapshot support.

3.15.2 Status

Deferred.