building_data_management_systems.Xuanzhou.2024Fall.DaSE
/
levelDB-LAB1-pzy_ckx

<!DOCTYPE html><html><head><link rel="stylesheet" type="text/css" href="doc.css" /><title>Leveldb file layout and compactions</title></head>
<body>
<h1>Files</h1>
The implementation of leveldb is similar in spirit to therepresentation of a single<a href="http://research.google.com/archive/bigtable.html">Bigtable tablet (section 5.3)</a>.However the organization of the files that make up the representationis somewhat different and is explained below.
<p>Each database is represented by a set of files stored in a directory.There are several different types of files as documented below:<p><h2>Log files</h2><p>A log file (*.log) stores a sequence of recent updates.  Each updateis appended to the current log file.  When the log file reaches apre-determined size (approximately 4MB by default), it is convertedto a sorted table (see below) and a new log file is created for futureupdates.<p>A copy of the current log file is kept in an in-memory structure (the<code>memtable</code>).  This copy is consulted on every read so that readoperations reflect all logged updates.<p><h2>Sorted tables</h2><p>A sorted table (*.sst) stores a sequence of entries sorted by key.Each entry is either a value for the key, or a deletion marker for thekey.  (Deletion markers are kept around to hide obsolete valuespresent in older sorted tables).<p>The set of sorted tables are organized into a sequence of levels.  Thesorted table generated from a log file is placed in a special <code>young</code>level (also called level-0).  When the number of young files exceeds acertain threshold (currently four), all of the young files are mergedtogether with all of the overlapping level-1 files to produce asequence of new level-1 files (we create a new level-1 file for every2MB of data.)<p>Files in the young level may contain overlapping keys.  However filesin other levels have distinct non-overlapping key ranges.  Considerlevel number L where L >= 1.  When the combined size of files inlevel-L exceeds (10^L) MB (i.e., 10MB for level-1, 100MB for level-2,...), one file in level-L, and all of the overlapping files inlevel-(L+1) are merged to form a set of new files for level-(L+1).These merges have the effect of gradually migrating new updates fromthe young level to the largest level using only bulk reads and writes(i.e., minimizing expensive seeks).
<h2>Manifest</h2><p>A MANIFEST file lists the set of sorted tables that make up eachlevel, the corresponding key ranges, and other important metadata.A new MANIFEST file (with a new number embedded in the file name)is created whenever the database is reopened.  The MANIFEST file isformatted as a log, and changes made to the serving state (as filesare added or removed) are appended to this log.<p><h2>Current</h2><p>CURRENT is a simple text file that contains the name of the latestMANIFEST file.<p><h2>Info logs</h2><p>Informational messages are printed to files named LOG and LOG.old.<p><h2>Others</h2><p>Other files used for miscellaneous purposes may also be present(LOCK, *.dbtmp).
<h1>Level 0</h1>When the log file grows above a certain size (1MB by default):<ul><li>Create a brand new memtable and log file and direct future updates here<li>In the background:<ul><li>Write the contents of the previous memtable to an sstable<li>Discard the memtable<li>Delete the old log file and the old memtable<li>Add the new sstable to the young (level-0) level.</ul></ul>
<h1>Compactions</h1>
<p>When the size of level L exceeds its limit, we compact it in abackground thread.  The compaction picks a file from level L and alloverlapping files from the next level L+1.  Note that if a level-Lfile overlaps only part of a level-(L+1) file, the entire file atlevel-(L+1) is used as an input to the compaction and will bediscarded after the compaction.  Aside: because level-0 is special(files in it may overlap each other), we treat compactions fromlevel-0 to level-1 specially: a level-0 compaction may pick more thanone level-0 file in case some of these files overlap each other.
<p>A compaction merges the contents of the picked files to produce asequence of level-(L+1) files.  We switch to producing a newlevel-(L+1) file after the current output file has reached the targetfile size (2MB).  We also switch to a new output file when the keyrange of the current output file has grown enough to overlap more thanten level-(L+2) files.  This last rule ensures that a later compactionof a level-(L+1) file will not pick up too much data from level-(L+2).
<p>The old files are discarded and the new files are added to the servingstate.
<p>Compactions for a particular level rotate through the key space.  Inmore detail, for each level L, we remember the ending key of the lastcompaction at level L.  The next compaction for level L will pick thefirst file that starts after this key (wrapping around to thebeginning of the key space if there is no such file).
<p>Compactions drop overwritten values.  They also drop deletion markersif there are no higher numbered levels that contain a file whose rangeoverlaps the current key.
<h2>Timing</h2>
Level-0 compactions will read up to four 1MB files from level-0, andat worst all the level-1 files (10MB).  I.e., we will read 14MB andwrite 14MB.
<p>Other than the special level-0 compactions, we will pick one 2MB filefrom level L.  In the worst case, this will overlap ~ 12 files fromlevel L+1 (10 because level-(L+1) is ten times the size of level-L,and another two at the boundaries since the file ranges at level-Lwill usually not be aligned with the file ranges at level-L+1).  Thecompaction will therefore read 26MB and write 26MB.  Assuming a diskIO rate of 100MB/s (ballpark range for modern drives), the worstcompaction cost will be approximately 0.5 second.
<p>If we throttle the background writing to something small, say 10% ofthe full 100MB/s speed, a compaction may take up to 5 seconds.  If theuser is writing at 10MB/s, we might build up lots of level-0 files(~50 to hold the 5*10MB).  This may significantly increase the cost ofreads due to the overhead of merging more files together on everyread.
<p>Solution 1: To reduce this problem, we might want to increase the logswitching threshold when the number of level-0 files is large.  Thoughthe downside is that the larger this threshold, the more memory we willneed to hold the corresponding memtable.
<p>Solution 2: We might want to decrease write rate artificially when thenumber of level-0 files goes up.
<p>Solution 3: We work on reducing the cost of very wide merges.Perhaps most of the level-0 files will have their blocks sittinguncompressed in the cache and we will only need to worry about theO(N) complexity in the merging iterator.
<h2>Number of files</h2>
Instead of always making 2MB files, we could make larger files forlarger levels to reduce the total file count, though at the expense ofmore bursty compactions.  Alternatively, we could shard the set offiles into multiple directories.
<p>An experiment on an <code>ext3</code> filesystem on Feb 04, 2011 showsthe following timings to do 100K file opens in directories withvarying number of files:<table class="datatable"><tr><th>Files in directory</th><th>Microseconds to open a file</th></tr><tr><td>1000</td><td>9</td><tr><td>10000</td><td>10</td><tr><td>100000</td><td>16</td></table>So maybe even the sharding is not necessary on modern filesystems?
<h1>Recovery</h1>
<ul><li> Read CURRENT to find name of the latest committed MANIFEST<li> Read the named MANIFEST file<li> Clean up stale files<li> We could open all sstables here, but it is probably better to be lazy...<li> Convert log chunk to a new level-0 sstable<li> Start directing new writes to a new log file with recovered sequence#</ul>
<h1>Garbage collection of files</h1>
<code>DeleteObsoleteFiles()</code> is called at the end of everycompaction and at the end of recovery.  It finds the names of allfiles in the database.  It deletes all log files that are not thecurrent log file.  It deletes all table files that are not referencedfrom some level and are not the output of an active compaction.
</body></html>