Ternary Search Tree for Prefix-Search
Copyright David C. Rankin, J.D.,P.E. 2017
Licensed Under GPLv2 available here
https://www.gnu.org/licenses/old-licenses/gpl-2.0.en.html
A ternary search tree (tst) is a binary search tree (bst) with an additional node pointer.
There is a relative dearth of information available about ternary trees, and specifically, proper node-rotation when a node is deleted from the tree. The only examples for a node-delete found are simple deletes that leave the tree dirty by leaving a node, with or without siblings, but having no valid middle (or equal) pointer. The following code, written as part of a text-editor word-completion implementation, provides a proper delete with rotation of the low or high node in place of the deleted node.
Ternary Search Tree Intro
With a binary search tree each node contains a left and right node-pointer so a binary choice controls traversal. Either a greater than or less than choice. As the result of a comparison between the node-data and a reference either the left or right node-pointer is used to further descend.
A ternary search tree adds a third (or middle) node. While you can still use the left-middle-right notation, ternary trees use a low-equal-high pointer verbiage. With each node holding a key ID, the node pointers for this code uses a lokid, eqkid, and hikid pointer naming convention. If the difference between a reference and the node key is negative (lower than), the lokid node is traversed, if they are equal, the eqkid node is followed or hikid is followed. The addition of the equal pointer and storage of character data as the node->key makes the ternary search tree optimal for string and prefix searching.
Tree-node Used In This Code
In addition to the node->key, the node used in this example adds a reference count (a node->refcnt) to the node data to track the number of occurrences for each word it holds. So for example, if using the tree to track the words in an editor buffer (where there may be multiple occurrences of 'the' or other common words), the node holding 'the' is not deleted, until no other occurrences remain (i.e. the node->refcnt is zero).
Each individual node has the following form:
o
|-key
|-refcnt
------------+------------
|lokid |eqkid |hikid
o o o
The string data (a pointer to a word, or a copy of the word itself) is stored as a pointer in an additional/special node following the node containing the last character (node->key) in the search path. The pointer to the data is cast to type (node *) and stored as the node->eqkid pointer. Further, since this is the node after the last character, similar to the end of a string, its key is the nul-character (decimal 0). The 'key's for each of the nodes that make up the search path of a word, are the letters in the word with the final node having a key 0 with either a pointer to the string (if stored in an external data structure) or an allocated copy of the string itself if the string is to be stored in the tree. (as in holding the words for an edit buffer, where the location/address for the string changes with each keypress). In either case, the traversal to the final node will have a form similar to the following for the word "cat":
o
|-c
|-0
------------+------------
|lokid |eqkid |hikid
o o o
|-a
|-0
----+----
| | | note: any of the lokid or hikid nodes
o can also have pointers to nodes
|-t for words that "cat" or "ca" is
|-0 a partial prefix to.
----+----
| | |
o
|-0
|-1 <== the refcnt is only relevant to the final node
----+----
| | |
NULL o NULL
"cat"
The ternary tree has the same O(n) efficiency for insert and search as does a bst. A dirty delete from the tree is O(n). This delete with rotation is only slightly less efficient due to the proper deletion of the chain of all unique nodes in the search path and proper rotation. Lookup times associated with loading the entire /usr/share/dict/words file and searching range between 0.000002 - 0.000014 sec. However, the prefix search ability offered by the ternary search tree sets it apart from virtually all other data structures. While Tri/Radix trees can perform as well, their memory requirements to cover all ASCII characters are often 20 times that of a ternary tree.
Example words File Provided (or use dict/words)
The /usr/share/dict/words file (or /var/lib/dict/words on some systems) holds roughly 305000 words (including plural possessives), and roughly 249092 words without. You can load the full words file for testing, or I have also provided the first 1000 words from the file in words1000.txt if you are not on a Linux/Unix machine. There is also a nice english-words list here on github with 370099 words (up to 31-chars, e.g. "dichlorodiphenyltrichloroethane") that can be used with the test programs provided along with this code. (note the words_alpha.txt file has DOS line endings). The test routines handle either DOS of POSIX line endings.
Ternary Search Tree Source Files
The files containing the ternary tree implementation are:
ternary_st.h
ternary_st.c
Interactive Example Programs
The following two test files provide a short menu driven application:
tst_test_cpy.c /* uses tst_ins_del_cpy() to store an allocated copy of word */
tst_test_ref.c /* uses tst_ins_del_ref() to store reference to word stored elsewhere */
The test application provides the following operations on the tree:
p print words in tree
a add word to the tree
f find word in tree
s search words matching prefix (enter 3 chars)
d delete word from the tree
q quit, freeing all data
(note: code skips printing tree with more than 100 words)
Tree Validation Program
The final file tst_validate.c is a short torture test fully exercising and validating tree integrity. After the tree is filled, the array of pointer to words used to fill the tree are shuffled into a random order and delete is called on each word in the shuffled list. A further inner loop than validates that every remaining word and every pointer in the tree each time a deletion takes place. It is a validation routine, so the only output provided is the following on success, on error, a dump of the deletion order file and current list state is made. The successful output is similar to:
$ tst_validate dat/dictwords_1000.txt
ternary_search_tree, loaded, 1000 words.
1000 successful deletions from search tree.
Compilation
Compilation with full error checking and optimization is suggested, e.g. and a Makefile is provided that will build the ternary_st.o object file and then compile all test programs placing the executables in a ./bin subdirectory. You can individually compile any of the test programs similar to the following:
$ gcc -Wall -Wextra -pedantic -Wshadow -finline-functions -std=c11 -Ofast \
ternary_st.c -o bin/ternary_st_val ternary_st_val.c
Prefix Searching
The benefit of a ternary tree for prefix searching of text lies in its ability to quickly traverse a tree of any size finding the node containing the last character in the wanted prefix. An in-order traversal of that node identifies all strings in the tree containing the prefix. For example, in the words1000.txt file, you can locate all words beginning with abr near instantaneously, e.g.
$ ./bin/tst_test_cpy words1000.txt
ternary_tree, loaded 1000 words in 0.001965 sec
p print words in tree
a add word to the tree
f find word in tree
s search words matching prefix (enter 3 chars)
d delete word from the tree
q quit, freeing all data
choice: s
find words matching prefix (3 chars): abr
abr - searched prefix in 0.000008 sec
suggest[0] : abrade
suggest[1] : abrasion
suggest[2] : abrasive
suggest[3] : abreact
suggest[4] : abreast
suggest[5] : abridge
suggest[6] : abridgment
suggest[7] : abroad
suggest[8] : abrogate
suggest[9] : abrupt
The search and prefix search times remain fast as the size of the tree increases. For example, adding all 249092 word from /usr/share/dict/words (not containing apostrophes), while there are eight times as many words with the abr prefix, the time to acquire an array of pointers to all prefixed words increases only slightly, e.g.
$ ./bin/tst_test_cpy dat/words
ternary_tree, loaded 249092 words in 0.186725 sec
p print words in tree
a add word to the tree
f find word in tree
s search words matching prefix (enter 3 chars)
d delete word from the tree
q quit, freeing all data
choice: s
find words matching prefix (3 chars): abr
abr - searched prefix in 0.000051 sec
suggest[0] : abrégé
suggest[1] : abr
suggest[2] : abracadabra
suggest[3] : abrachia
suggest[4] : abradable
suggest[5] : abradant
suggest[6] : abrade
<snip>
suggest[67] : abrupter
suggest[68] : abruptest
suggest[69] : abruption
suggest[70] : abruptly
suggest[71] : abruptness
suggest[72] : abruptnesses
Changes
With the addition of a Makefile the source tree has been reorganized to separate the source and include files into separate directories as well as moving the example file to the dat subdirectory. The source tree layout is now:
├── Makefile
├── Makefile.lib
├── README.md
├── dat
│ └── words1000.txt
├── include
│ └── ternary_st.h
└── src
├── ternary_st.c
├── tst_test_cpy.c
├── tst_test_ref.c
└── tst_validate.c
strdup has been removed and replaced with standard strlen, malloc, memcpy to address the removal of strdup from the C99 standard. While many compilers provide strdup as an extension, as was originally provided by -std=gnu11, the code is now C11 standard compliant and backward compatible with prior versions.
An additional makefile Makefile.lib was added which builds a Linux shared-object library from the ternary_st source. The original makefile builds the .so by default by calling Makefile.lib within the all: rule. You can build-only the shared-object library by invoking make as make libternary_st or by simply passing the -f option to call make on the library makefile with make -f Makefile.lib.
Adding install (and running with elevated privileges) will install the shared-object library in $(PREFIX)/$(INSTDIR) (/usr/local/lib64 by default). The configuration file containing the path is installed in /etc/ld.so.conf.d/ternary_st.conf and ldconfig is run to update the linker shared-object library cache.
If you find any problems, let me know (or better yet, create a pull-request and provide a patch).