diff --git a/src/backend/access/common/heaptuple.c b/src/backend/access/common/heaptuple.c index aea9d40..5619825 100644 --- a/src/backend/access/common/heaptuple.c +++ b/src/backend/access/common/heaptuple.c @@ -60,6 +60,10 @@ #include "access/sysattr.h" #include "access/tuptoaster.h" #include "executor/tuptable.h" +#include "utils/pg_lzcompress.h" + +/* guc variable for EWT compression ratio. */ +int wal_update_compression_ratio = 25; /* Does att's datatype allow packing into the 1-byte-header varlena format? */ @@ -617,6 +621,49 @@ heap_copytuple_with_tuple(HeapTuple src, HeapTuple dest) memcpy((char *) dest->t_data, (char *) src->t_data, src->t_len); } +/* ---------------- + * heap_delta_encode + * + * Calculate the delta between two tuples and generate + * encoded wal tuple (EWT), using pgrb. The result is stored + * in *encdata. + * ---------------- + */ +bool +heap_delta_encode(TupleDesc tupleDesc, HeapTuple oldtup, HeapTuple newtup, + char *encdata, uint32 *enclen) +{ + PGLZ_Strategy strategy; + + strategy = *PGLZ_strategy_default; + strategy.min_comp_rate = wal_update_compression_ratio; + + return pgrb_delta_encode( + (char *) newtup->t_data + offsetof(HeapTupleHeaderData, t_bits), + newtup->t_len - offsetof(HeapTupleHeaderData, t_bits), + (char *) oldtup->t_data + offsetof(HeapTupleHeaderData, t_bits), + oldtup->t_len - offsetof(HeapTupleHeaderData, t_bits), + encdata, enclen, &strategy + ); +} + +/* ---------------- + * heap_delta_decode + * + * Decode a tuple using delta-encoded WAL tuple and old tuple version. + * ---------------- + */ +void +heap_delta_decode(char *encdata, uint32 enclen, HeapTuple oldtup, HeapTuple newtup) +{ + pgrb_delta_decode(encdata, enclen, + (char *) newtup->t_data + offsetof(HeapTupleHeaderData, t_bits), + MaxHeapTupleSize - offsetof(HeapTupleHeaderData, t_bits), + &newtup->t_len, + (char *) oldtup->t_data + offsetof(HeapTupleHeaderData, t_bits), + oldtup->t_len - offsetof(HeapTupleHeaderData, t_bits)); +} + /* * heap_form_tuple * construct a tuple from the given values[] and isnull[] arrays, diff --git a/src/backend/access/common/reloptions.c b/src/backend/access/common/reloptions.c index fa08c45..0e9efca 100644 --- a/src/backend/access/common/reloptions.c +++ b/src/backend/access/common/reloptions.c @@ -85,6 +85,14 @@ static relopt_bool boolRelOpts[] = }, false }, + { + { + "compress_wal", + "Compress the wal tuple for update operation on this relation", + RELOPT_KIND_HEAP + }, + true + }, /* list terminator */ {{NULL}} }; @@ -1175,7 +1183,9 @@ default_reloptions(Datum reloptions, bool validate, relopt_kind kind) {"check_option", RELOPT_TYPE_STRING, offsetof(StdRdOptions, check_option_offset)}, {"user_catalog_table", RELOPT_TYPE_BOOL, - offsetof(StdRdOptions, user_catalog_table)} + offsetof(StdRdOptions, user_catalog_table)}, + {"compress_wal", RELOPT_TYPE_BOOL, + offsetof(StdRdOptions, compress_wal)} }; options = parseRelOptions(reloptions, validate, kind, &numoptions); diff --git a/src/backend/access/heap/heapam.c b/src/backend/access/heap/heapam.c index 05c790f..cf6f1fb 100644 --- a/src/backend/access/heap/heapam.c +++ b/src/backend/access/heap/heapam.c @@ -70,6 +70,7 @@ #include "utils/snapmgr.h" #include "utils/syscache.h" #include "utils/tqual.h" +#include "utils/pg_lzcompress.h" /* GUC variable */ @@ -6591,6 +6592,12 @@ log_heap_update(Relation reln, Buffer oldbuf, XLogRecPtr recptr; XLogRecData rdata[7]; Page page = BufferGetPage(newbuf); + char *newtupdata; + int newtuplen; + bool compressed = false; + + /* Structure which holds EWT */ + char buf[MaxHeapTupleSize]; bool need_tuple_data = RelationIsLogicallyLogged(reln); /* Caller should not call me on a non-WAL-logged relation */ @@ -6601,6 +6608,38 @@ log_heap_update(Relation reln, Buffer oldbuf, else info = XLOG_HEAP_UPDATE; + newtupdata = ((char *) newtup->t_data) + offsetof(HeapTupleHeaderData, t_bits); + newtuplen = newtup->t_len - offsetof(HeapTupleHeaderData, t_bits); + + /* + * EWT can be generated for all new tuple versions created by Update + * operation. Currently we do it when both the old and new tuple versions + * are on same page, because during recovery if the page containing old + * tuple is corrupt, it should not cascade that corruption to other pages. + * Under the general assumption that for long runs most updates tend to + * create new tuple version on same page, there should not be significant + * impact on WAL reduction or performance. + * + * We should not generate EWT when we need to backup the whole block in + * WAL as in that case there is no saving by reduced WAL size. + */ + + if (RelationIsEnabledForWalCompression(reln) && + wal_update_compression_ratio != 0 && + (oldbuf == newbuf) && + !XLogCheckBufferNeedsBackup(newbuf)) + { + uint32 enclen; + + /* Delta-encode the new tuple using the old tuple */ + if (heap_delta_encode(reln->rd_att, oldtup, newtup, buf, &enclen)) + { + compressed = true; + newtupdata = buf; + newtuplen = enclen; + } + } + xlrec.target.node = reln->rd_node; xlrec.target.tid = oldtup->t_self; xlrec.old_xmax = HeapTupleHeaderGetRawXmax(oldtup->t_data); @@ -6613,6 +6652,8 @@ log_heap_update(Relation reln, Buffer oldbuf, xlrec.newtid = newtup->t_self; if (new_all_visible_cleared) xlrec.flags |= XLOG_HEAP_NEW_ALL_VISIBLE_CLEARED; + if (compressed) + xlrec.flags |= XLOG_HEAP_DELTA_ENCODED; rdata[0].data = (char *) &xlrec; rdata[0].len = SizeOfHeapUpdate; @@ -6628,7 +6669,7 @@ log_heap_update(Relation reln, Buffer oldbuf, xlhdr.header.t_infomask2 = newtup->t_data->t_infomask2; xlhdr.header.t_infomask = newtup->t_data->t_infomask; xlhdr.header.t_hoff = newtup->t_data->t_hoff; - xlhdr.t_len = newtup->t_len - offsetof(HeapTupleHeaderData, t_bits); + xlhdr.t_len = newtuplen; /* * As with insert records, we need not store the rdata[2] segment @@ -6641,10 +6682,13 @@ log_heap_update(Relation reln, Buffer oldbuf, rdata[2].buffer_std = true; rdata[2].next = &(rdata[3]); - /* PG73FORMAT: write bitmap [+ padding] [+ oid] + data */ - rdata[3].data = (char *) newtup->t_data - + offsetof(HeapTupleHeaderData, t_bits); - rdata[3].len = newtup->t_len - offsetof(HeapTupleHeaderData, t_bits); + /* + * PG73FORMAT: write bitmap [+ padding] [+ oid] + data OR + * PG94FORMAT [If encoded]: Control byte + history reference (2 - 3)bytes + * + literal byte + ... + */ + rdata[3].data = newtupdata; + rdata[3].len = newtuplen; rdata[3].buffer = need_tuple_data ? InvalidBuffer : newbuf; rdata[3].buffer_std = true; rdata[3].next = NULL; @@ -7733,7 +7777,10 @@ heap_xlog_update(XLogRecPtr lsn, XLogRecord *record, bool hot_update) Page page; OffsetNumber offnum; ItemId lp = NULL; + HeapTupleData newtup; + HeapTupleData oldtup; HeapTupleHeader htup; + HeapTupleHeader oldtupdata = NULL; struct { HeapTupleHeaderData hdr; @@ -7808,7 +7855,7 @@ heap_xlog_update(XLogRecPtr lsn, XLogRecord *record, bool hot_update) if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp)) elog(PANIC, "heap_update_redo: invalid lp"); - htup = (HeapTupleHeader) PageGetItem(page, lp); + oldtupdata = htup = (HeapTupleHeader) PageGetItem(page, lp); htup->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED); htup->t_infomask2 &= ~HEAP_KEYS_UPDATED; @@ -7917,10 +7964,31 @@ newsame:; Assert(newlen <= MaxHeapTupleSize); htup = &tbuf.hdr; MemSet((char *) htup, 0, sizeof(HeapTupleHeaderData)); - /* PG73FORMAT: get bitmap [+ padding] [+ oid] + data */ - memcpy((char *) htup + offsetof(HeapTupleHeaderData, t_bits), - (char *) xlrec + hsize, - newlen); + + /* + * If the record is EWT then decode it. + */ + if (xlrec->flags & XLOG_HEAP_DELTA_ENCODED) + { + /* + * PG94FORMAT: Control byte + history reference (2 - 3)bytes + * + literal byte + ... + */ + oldtup.t_data = oldtupdata; + oldtup.t_len = ItemIdGetLength(lp); + newtup.t_data = htup; + + heap_delta_decode((char *) xlrec + hsize, newlen, &oldtup, &newtup); + newlen = newtup.t_len; + } + else + { + /* PG73FORMAT: get bitmap [+ padding] [+ oid] + data */ + memcpy((char *) htup + offsetof(HeapTupleHeaderData, t_bits), + (char *) xlrec + hsize, + newlen); + } + newlen += offsetof(HeapTupleHeaderData, t_bits); htup->t_infomask2 = xlhdr.header.t_infomask2; htup->t_infomask = xlhdr.header.t_infomask; diff --git a/src/backend/access/transam/xlog.c b/src/backend/access/transam/xlog.c index 3cde91e..05570bd 100644 --- a/src/backend/access/transam/xlog.c +++ b/src/backend/access/transam/xlog.c @@ -2324,6 +2324,28 @@ XLogRecPtrToBytePos(XLogRecPtr ptr) } /* + * Determine whether the buffer referenced has to be backed up. Since we don't + * yet have the insert lock, fullPageWrites and forcePageWrites could change + * later, but will not cause any problem because this function is used only to + * identify whether EWT is required for update. + */ +bool +XLogCheckBufferNeedsBackup(Buffer buffer) +{ + bool doPageWrites; + Page page; + + page = BufferGetPage(buffer); + + doPageWrites = XLogCtl->Insert.fullPageWrites || XLogCtl->Insert.forcePageWrites; + + if (doPageWrites && PageGetLSN(page) <= RedoRecPtr) + return true; /* buffer requires backup */ + + return false; /* buffer does not need to be backed up */ +} + +/* * Determine whether the buffer referenced by an XLogRecData item has to * be backed up, and if so fill a BkpBlock struct for it. In any case * save the buffer's LSN at *lsn. diff --git a/src/backend/utils/adt/pg_lzcompress.c b/src/backend/utils/adt/pg_lzcompress.c index 30f1c0a..ab90940 100644 --- a/src/backend/utils/adt/pg_lzcompress.c +++ b/src/backend/utils/adt/pg_lzcompress.c @@ -186,6 +186,22 @@ #define PGLZ_MAX_HISTORY_LISTS 8192 /* must be power of 2 */ #define PGLZ_HISTORY_SIZE 4096 #define PGLZ_MAX_MATCH 273 +/* + * Popular and effective rolling hash function treats every substring + * as a number in some base, the base being usually a prime. + * Experiments suggest that prime number 11 generates better chunks. + * Currently experiements have been done on repetitive data, more + * experiments can be done with different kind of data to choose + * more appropriate prime number. + */ +#define PGRB_HKEY_PRIME 11 /* prime number used for rolling hash */ +#define PGRB_HKEY_SQUARE_PRIME 11 * 11 /* square of prime number used for rolling hash */ +#define PGRB_HKEY_CUBE_PRIME 11 * 11 * 11 /* cube of prime number used for rolling hash */ +/* number of bits after which to check for constant pattern to form chunk */ +#define PGRB_PATTERN_AFTER_BITS 4 +#define PGRB_CONST_NUM (1 << PGRB_PATTERN_AFTER_BITS) +#define PGRB_MIN_CHUNK_SIZE 2 +#define PGRB_MAX_CHUNK_SIZE 4 /* ---------- @@ -239,12 +255,21 @@ static const PGLZ_Strategy strategy_always_data = { const PGLZ_Strategy *const PGLZ_strategy_always = &strategy_always_data; +typedef struct PGRB_HistEntry +{ + struct PGRB_HistEntry *next; /* links for my hash key's list */ + uint32 hindex; /* my current hash key */ + const char *ck_start_pos; /* chunk start position */ + int16 ck_size; /* chunk end position */ +} PGRB_HistEntry; + /* ---------- * Statically allocated work arrays for history * ---------- */ static int16 hist_start[PGLZ_MAX_HISTORY_LISTS]; static PGLZ_HistEntry hist_entries[PGLZ_HISTORY_SIZE + 1]; +static PGRB_HistEntry rb_hist_entries[PGLZ_HISTORY_SIZE + 1]; /* * Element 0 in hist_entries is unused, and means 'invalid'. Likewise, @@ -252,6 +277,7 @@ static PGLZ_HistEntry hist_entries[PGLZ_HISTORY_SIZE + 1]; */ #define INVALID_ENTRY 0 #define INVALID_ENTRY_PTR (&hist_entries[INVALID_ENTRY]) +#define RB_INVALID_ENTRY_PTR (&rb_hist_entries[INVALID_ENTRY]) /* ---------- * pglz_hist_idx - @@ -271,6 +297,49 @@ static PGLZ_HistEntry hist_entries[PGLZ_HISTORY_SIZE + 1]; ((_s)[2] << 2) ^ (_s)[3])) & (_mask) \ ) +/* + * pgrb_hash_init and pgrb_hash_roll can be use to calculate the hash in + * a rolling fashion. First, call pgrb_hash_init, with a pointer to the first + * byte. Then call pglz_hash_roll for every subsequent byte. After each + * pglz_hash_roll() call, hindex holds the hash of the current byte. + * + * a,b,c,d are local variables these macros use to store state. These macros + * don't check for end-of-buffer like pglz_hist_idx() does, so these cannot be + * used on the last 3 bytes of input. + */ +#define pgrb_hash_init(_p,hindex,a,b,c,d) \ + do { \ + a = _p[0]; \ + b = _p[1]; \ + c = _p[2]; \ + d = _p[3]; \ + hindex = (a * PGRB_HKEY_CUBE_PRIME + b * PGRB_HKEY_SQUARE_PRIME + c * PGRB_HKEY_PRIME + d); \ + } while (0) + +#define pgrb_hash_roll(_p,hindex,a,b,c,d) \ + do { \ + /* subtract old a, 1000 % 11 = 10 */ \ + hindex -= (a * PGRB_HKEY_CUBE_PRIME); \ + /* add new byte */ \ + a = b; b = c; c = d; d = _p[3]; \ + hindex = PGRB_HKEY_PRIME * hindex + d; \ + } while (0) + +/* + * An version of pglz_hist_add() that doesn't do recycling. Can be used if + * you know the input fits in PGLZ_HISTORY_SIZE. + */ +#define pgrb_hist_add_no_recycle(_hs,_he,_hn,_s,_ck_size, _hindex) \ +do { \ + int16 *__myhsp = &(_hs)[_hindex]; \ + PGRB_HistEntry *__myhe = &(_he)[_hn]; \ + __myhe->next = &(_he)[*__myhsp]; \ + __myhe->hindex = _hindex; \ + __myhe->ck_start_pos = (_s); \ + __myhe->ck_size = (_ck_size); \ + *__myhsp = _hn; \ + ++(_hn); \ +} while (0) /* ---------- * pglz_hist_add - @@ -488,6 +557,107 @@ pglz_find_match(int16 *hstart, const char *input, const char *end, return 0; } +/* ---------- + * pgrb_find_match - + * + * ---------- + */ +static inline int +pgrb_find_match(int16 *hstart, const char* input_chunk_start, + int input_chunk_size, int16 *lenp, int16 *offp, + const char *hend, int hindex) +{ + PGRB_HistEntry *hent; + int16 hentno; + const char *hp; + const char *ip; + int16 history_chunk_size; + int16 matchlen; + bool match_chunk; + + hentno = hstart[hindex]; + hent = &rb_hist_entries[hentno]; + + while (hent != RB_INVALID_ENTRY_PTR) + { + /* + * if history and input chunk size doesn't match, then chunks cannot + * match. + */ + history_chunk_size = hent->ck_size; + if (history_chunk_size != input_chunk_size) + return 0; + + match_chunk = true; + matchlen = history_chunk_size; + hp = hent->ck_start_pos; + ip = input_chunk_start; + + /* + * first try to match uptil chunksize and if the data is + * same for chunk, then try to match further to get the + * larger match. if there is a match at end of chunk, it + * is possible that further bytes in string will match. + */ + while (history_chunk_size-- > 0) + { + if (*hp++ != *ip++) + { + match_chunk = false; + break; + } + else + match_chunk = true; + } + + if (match_chunk) + { + while (*ip == *hp) + { + matchlen++; + ip++; + hp++; + } + } + else + { + hent = hent->next; + continue; + } + + *offp = hend - hent->ck_start_pos; + *lenp = matchlen; + + return 1; + } + + return 0; +} + +static int +choose_hash_size(int slen) +{ + int hashsz; + + /* + * Experiments suggest that these hash sizes work pretty well. A large + * hash table minimizes collision, but has a higher startup cost. For a + * small input, the startup cost dominates. The table size must be a power + * of two. + */ + if (slen < 128) + hashsz = 512; + else if (slen < 256) + hashsz = 1024; + else if (slen < 512) + hashsz = 2048; + else if (slen < 1024) + hashsz = 4096; + else + hashsz = 8192; + return hashsz; +} + /* ---------- * pglz_compress - @@ -676,6 +846,259 @@ pglz_compress(const char *source, int32 slen, PGLZ_Header *dest, /* ---------- + * Rabin's Delta Encoding - + * + * The 'source' is encoded using rabin finger print method. + * We use a rolling hash function to divide up the history data into chunks + * of a given average size. To do that we scan the history data, compute a + * rolling hash value at each byte, and each time the bottom + * PGRB_PATTERN_AFTER_BITS bits are zero, we consider that to be the end of + * a chunk. To make the chunk size more predictable and handle worst case + * (the data doesn't contain special pattern) we use min and max chunk + * boundaries. Enter all the chunks into a hash table. + * Then, we scan the input we want to compress and divide it into chunks in + * the same way. Chunks that don't exist in the history data get copied to + * the output, while those that do get replaced with a reference to their + * position in the history data. + * The encoding format to store encoded string is same as pglz. + * + * The encoded result is written to *dest, and its length is returned in + * *finallen. + */ +bool +pgrb_delta_encode(const char *source, int32 slen, + const char *history, int32 hlen, + char *dest, uint32 *finallen, + const PGLZ_Strategy *strategy) +{ + unsigned char *bp = ((unsigned char *) dest); + unsigned char *bstart = bp; + const char *dp = source; + const char *dp_chunk_start; + const char *dend = source + slen; + const char *dp_unmatched_chunk_start; + const char *hp = history; + const char *hp_chunk_start; + const char *hend = history + hlen; + unsigned char ctrl_dummy = 0; + unsigned char *ctrlp = &ctrl_dummy; + unsigned char ctrlb = 0; + unsigned char ctrl = 0; + bool found_match = false; + bool unmatched_data = false; + int16 match_len = 0; + int16 match_off; + int32 result_size; + int32 result_max; + int32 need_rate; + int hist_next = 1; + int hashsz; + int mask; + int32 a, + b, + c, + d; + int32 hindex; + int16 len = 0; + int16 dp_chunk_size = 1; + int16 hp_chunk_size = 1; + + /* + * Tuples of length greater than PGLZ_HISTORY_SIZE are not allowed for + * delta encode as this is the maximum size of history offset. + */ + if (hlen >= PGLZ_HISTORY_SIZE || hlen < 4) + return false; + + /* + * Our fallback strategy is the default. + */ + if (strategy == NULL) + strategy = PGLZ_strategy_default; + + /* + * If the strategy forbids compression (at all or if source chunk size out + * of range), fail. + */ + if (strategy->match_size_good <= 0 || + slen < strategy->min_input_size || + slen > strategy->max_input_size) + return false; + + need_rate = strategy->min_comp_rate; + if (need_rate < 0) + need_rate = 0; + else if (need_rate > 99) + need_rate = 99; + + + /* + * Compute the maximum result size allowed by the strategy, namely the + * input size minus the minimum wanted compression rate. This had better + * be <= slen, else we might overrun the provided output buffer. + */ + if (slen > (INT_MAX / 100)) + { + /* Approximate to avoid overflow */ + result_max = (slen / 100) * (100 - need_rate); + } + else + result_max = (slen * (100 - need_rate)) / 100; + + hashsz = choose_hash_size(hlen/PGRB_MIN_CHUNK_SIZE); + mask = hashsz - 1; + + /* + * Initialize the history lists to empty. We do not need to zero the + * hist_entries[] array; its entries are initialized as they are used. + */ + memset(hist_start, 0, hashsz * sizeof(int16)); + + hp_chunk_start = hp; + pgrb_hash_init(hp, hindex, a, b, c, d); + while (hp < hend - 4) + { + /* + * if we found the special pattern or reached max chunk size, + * then consider it as a chunk and add the same to history + * table. + */ + if ((hp_chunk_size >= PGRB_MIN_CHUNK_SIZE && + (hindex % PGRB_CONST_NUM) == 0) || + hp_chunk_size == PGRB_MAX_CHUNK_SIZE) + { + pgrb_hist_add_no_recycle(hist_start, rb_hist_entries, + hist_next, hp_chunk_start, + hp_chunk_size, (hindex & mask)); + hp++; + hp_chunk_start = hp; + hp_chunk_size = 1; + } + else + { + hp++; /* Do not do this ++ in the line above! */ + hp_chunk_size++; + } + pgrb_hash_roll(hp, hindex, a, b, c, d); + } + + /* + * Loop through the input. + */ + match_off = 0; + dp_chunk_start = dp; + dp_unmatched_chunk_start = dp; + pgrb_hash_init(dp, hindex, a, b, c, d); + while (dp < dend - 4) + { + /* + * If we already exceeded the maximum result size, fail. + * + * We check once per loop; since the loop body could emit as many as 4 + * bytes (a control byte and 3-byte tag), PGLZ_MAX_OUTPUT() had better + * allow 4 slop bytes. + */ + if (found_match) + { + if (bp - bstart >= result_max) + return false; + } + else + { + if (len >= result_max) + return false; + } + + /* + * Try to find a match in the history + */ + if ((dp_chunk_size >= PGRB_MIN_CHUNK_SIZE && + (hindex % PGRB_CONST_NUM) == 0) || + dp_chunk_size == PGRB_MAX_CHUNK_SIZE) + { + if (pgrb_find_match(hist_start, dp_chunk_start, + dp_chunk_size, &match_len, &match_off, + hend, (hindex & mask))) + { + /* + * Create the tag and add history entries for all matched + * characters and ensure to copy any unmatched data till + * this point. Currently this code only delays copy of + * unmatched data in begining. + */ + if (unmatched_data) + { + while (dp_unmatched_chunk_start <= dp) + { + pglz_out_literal(ctrlp, ctrlb, ctrl, bp, *dp_unmatched_chunk_start); + dp_unmatched_chunk_start++; + } + unmatched_data = false; + } + pglz_out_tag(ctrlp, ctrlb, ctrl, bp, match_len, match_off); + found_match = true; + if (match_len > dp_chunk_size) + dp += match_len - dp_chunk_size; + } + else + { + /* No match found, copy chunk into destination buffer. */ + if (found_match) + { + while (dp_chunk_start <= dp) + { + pglz_out_literal(ctrlp, ctrlb, ctrl, bp, *dp_chunk_start); + dp_chunk_start++; + } + /* The macro would do it four times - Jan. */ + } + else + unmatched_data = true; + } + len++; + dp++; /* Do not do this ++ in the line above! */ + dp_chunk_start = dp; + dp_chunk_size = 1; + } + else + { + dp_chunk_size++; + len++; + dp++; + } + pgrb_hash_roll(dp, hindex, a, b, c, d); + } + + if (!found_match) + return false; + + /* Handle the last few bytes as literals */ + while (dp < dend) + { + pglz_out_literal(ctrlp, ctrlb, ctrl, bp, *dp); + dp++; /* Do not do this ++ in the line above! */ + } + + /* + * Write out the last control byte and check that we haven't overrun the + * output size allowed by the strategy. + */ + *ctrlp = ctrlb; + result_size = bp - bstart; + +#ifdef DELTA_DEBUG + elog(LOG, "old %d new %d compressed %d", hlen, slen, result_size); +#endif + + /* + * Success - need only fill in the actual length of the compressed datum. + */ + *finallen = result_size; + + return true; +} + +/* ---------- * pglz_decompress - * * Decompresses source into dest. @@ -777,3 +1200,107 @@ pglz_decompress(const PGLZ_Header *source, char *dest) * That's it. */ } + +/* ---------- + * pgrb_delta_decode + * + * Decompresses source into dest. + * To decompress, it uses history if provided. + * ---------- + */ +void +pgrb_delta_decode(const char *source, uint32 srclen, + char *dest, uint32 destlen, uint32 *finallen, + const char *history, uint32 histlen) +{ + const unsigned char *sp; + const unsigned char *srcend; + unsigned char *dp; + unsigned char *destend; + const char *hend; + + sp = ((const unsigned char *) source); + srcend = ((const unsigned char *) source) + srclen; + dp = (unsigned char *) dest; + destend = dp + destlen; + hend = history + histlen; + + while (sp < srcend && dp < destend) + { + /* + * Read one control byte and process the next 8 items (or as many as + * remain in the compressed input). + */ + unsigned char ctrl = *sp++; + int ctrlc; + + for (ctrlc = 0; ctrlc < 8 && sp < srcend; ctrlc++) + { + if (ctrl & 1) + { + /* + * Otherwise it contains the match length minus 3 and the + * upper 4 bits of the offset. The next following byte + * contains the lower 8 bits of the offset. If the length is + * coded as 18, another extension tag byte tells how much + * longer the match really was (0-255). + */ + int32 len; + int32 off; + + len = (sp[0] & 0x0f) + 3; + off = ((sp[0] & 0xf0) << 4) | sp[1]; + sp += 2; + if (len == 18) + len += *sp++; + + /* + * Check for output buffer overrun, to ensure we don't clobber + * memory in case of corrupt input. Note: we must advance dp + * here to ensure the error is detected below the loop. We + * don't simply put the elog inside the loop since that will + * probably interfere with optimization. + */ + if (dp + len > destend) + { + dp += len; + break; + } + + /* + * Now we copy the bytes specified by the tag from history to + * OUTPUT. + */ + memcpy(dp, hend - off, len); + dp += len; + } + else + { + /* + * An unset control bit means LITERAL BYTE. So we just copy + * one from INPUT to OUTPUT. + */ + if (dp >= destend) /* check for buffer overrun */ + break; /* do not clobber memory */ + + *dp++ = *sp++; + } + + /* + * Advance the control bit + */ + ctrl >>= 1; + } + } + + /* + * Check we decompressed the right amount. + */ + if (sp != srcend) + elog(PANIC, "compressed data is corrupt"); + + /* + * That's it. + */ + *finallen = ((char *) dp - dest); +} diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c index 1217098..d58b24d 100644 --- a/src/backend/utils/misc/guc.c +++ b/src/backend/utils/misc/guc.c @@ -2486,6 +2486,17 @@ static struct config_int ConfigureNamesInt[] = NULL, NULL, NULL }, + { + /* Not for general use */ + {"wal_update_compression_ratio", PGC_USERSET, DEVELOPER_OPTIONS, + gettext_noop("Sets the compression ratio of delta record for wal update"), + NULL, + }, + &wal_update_compression_ratio, + 25, 0, 100, + NULL, NULL, NULL + }, + /* End-of-list marker */ { {NULL, 0, 0, NULL, NULL}, NULL, 0, 0, 0, NULL, NULL, NULL diff --git a/src/include/access/heapam_xlog.h b/src/include/access/heapam_xlog.h index d4383ab..df64096 100644 --- a/src/include/access/heapam_xlog.h +++ b/src/include/access/heapam_xlog.h @@ -67,6 +67,7 @@ #define XLOG_HEAP_CONTAINS_OLD_TUPLE (1<<2) #define XLOG_HEAP_CONTAINS_OLD_KEY (1<<3) #define XLOG_HEAP_CONTAINS_NEW_TUPLE (1<<4) +#define XLOG_HEAP_DELTA_ENCODED (1<<5) /* convenience macro for checking whether any form of old tuple was logged */ #define XLOG_HEAP_CONTAINS_OLD \ diff --git a/src/include/access/htup_details.h b/src/include/access/htup_details.h index a3eba98..abb5620 100644 --- a/src/include/access/htup_details.h +++ b/src/include/access/htup_details.h @@ -740,6 +740,11 @@ extern HeapTuple heap_modify_tuple(HeapTuple tuple, extern void heap_deform_tuple(HeapTuple tuple, TupleDesc tupleDesc, Datum *values, bool *isnull); +extern bool heap_delta_encode(TupleDesc tupleDesc, HeapTuple oldtup, + HeapTuple newtup, char *encdata, uint32 *enclen); +extern void heap_delta_decode (char *encdata, uint32 enclen, HeapTuple oldtup, + HeapTuple newtup); + /* these three are deprecated versions of the three above: */ extern HeapTuple heap_formtuple(TupleDesc tupleDescriptor, Datum *values, char *nulls); diff --git a/src/include/access/xlog.h b/src/include/access/xlog.h index 017e74d..0aae944 100644 --- a/src/include/access/xlog.h +++ b/src/include/access/xlog.h @@ -278,6 +278,7 @@ typedef struct CheckpointStatsData extern CheckpointStatsData CheckpointStats; extern XLogRecPtr XLogInsert(RmgrId rmid, uint8 info, XLogRecData *rdata); +extern bool XLogCheckBufferNeedsBackup(Buffer buffer); extern void XLogFlush(XLogRecPtr RecPtr); extern bool XLogBackgroundFlush(void); extern bool XLogNeedsFlush(XLogRecPtr RecPtr); diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h index b145a19..eaffe64 100644 --- a/src/include/miscadmin.h +++ b/src/include/miscadmin.h @@ -226,6 +226,8 @@ extern bool allowSystemTableMods; extern PGDLLIMPORT int work_mem; extern PGDLLIMPORT int maintenance_work_mem; +extern int wal_update_compression_ratio; + extern int VacuumCostPageHit; extern int VacuumCostPageMiss; extern int VacuumCostPageDirty; diff --git a/src/include/utils/pg_lzcompress.h b/src/include/utils/pg_lzcompress.h index 4af24a3..3797b6c 100644 --- a/src/include/utils/pg_lzcompress.h +++ b/src/include/utils/pg_lzcompress.h @@ -107,6 +107,15 @@ extern const PGLZ_Strategy *const PGLZ_strategy_always; */ extern bool pglz_compress(const char *source, int32 slen, PGLZ_Header *dest, const PGLZ_Strategy *strategy); +extern bool pglz_delta_encode(const char *source, int32 slen, + const char *history, int32 hlen, + char *dest, uint32 *finallen, const PGLZ_Strategy *strategy); +extern bool pgrb_delta_encode(const char *source, int32 slen, + const char *history, int32 hlen, + char *dest, uint32 *finallen, const PGLZ_Strategy *strategy); extern void pglz_decompress(const PGLZ_Header *source, char *dest); +extern void pgrb_delta_decode(const char *source, uint32 srclen, + char *dest, uint32 destlen, uint32 *finallen, + const char *history, uint32 histlen); #endif /* _PG_LZCOMPRESS_H_ */ diff --git a/src/include/utils/rel.h b/src/include/utils/rel.h index 9b8a4c9..4c510ba 100644 --- a/src/include/utils/rel.h +++ b/src/include/utils/rel.h @@ -218,6 +218,7 @@ typedef struct StdRdOptions bool security_barrier; /* for views */ int check_option_offset; /* for views */ bool user_catalog_table; /* use as an additional catalog relation */ + bool compress_wal; /* compress wal tuple update */ } StdRdOptions; #define HEAP_MIN_FILLFACTOR 10 @@ -296,6 +297,15 @@ typedef struct StdRdOptions ((StdRdOptions *) (relation)->rd_options)->user_catalog_table : false) /* + * RelationIsEnabledForWalCompression + * Returns whether the wal for update operation on relation can + * be compressed. + */ +#define RelationIsEnabledForWalCompression(relation) \ + ((relation)->rd_options ? \ + ((StdRdOptions *) (relation)->rd_options)->compress_wal : true) + +/* * RelationIsValid * True iff relation descriptor is valid. */ diff --git a/src/test/regress/expected/update.out b/src/test/regress/expected/update.out index 71b856f..af46df2 100644 --- a/src/test/regress/expected/update.out +++ b/src/test/regress/expected/update.out @@ -97,3 +97,73 @@ SELECT a, b, char_length(c) FROM update_test; (2 rows) DROP TABLE update_test; +-- +-- Test to update continuos and non continuos columns +-- +DROP TABLE IF EXISTS update_test; +NOTICE: table "update_test" does not exist, skipping +CREATE TABLE update_test ( + bser bigserial, + bln boolean, + ename VARCHAR(25), + perf_f float(8), + grade CHAR, + dept CHAR(5) NOT NULL, + dob DATE, + idnum INT, + addr VARCHAR(30) NOT NULL, + destn CHAR(6), + Gend CHAR, + samba BIGINT, + hgt float, + ctime TIME +); +INSERT INTO update_test VALUES ( + nextval('update_test_bser_seq'::regclass), + TRUE, + 'Test', + 7.169, + 'B', + 'CSD', + '2000-01-01', + 520, + 'road2, + streeeeet2, + city2', + 'dcy2', + 'M', + 12000, + 50.4, + '00:00:00.0' +); +SELECT * from update_test; + bser | bln | ename | perf_f | grade | dept | dob | idnum | addr | destn | gend | samba | hgt | ctime +------+-----+-------+--------+-------+-------+------------+-------+-----------------------------+--------+------+-------+------+---------- + 1 | t | Test | 7.169 | B | CSD | 01-01-2000 | 520 | road2, +| dcy2 | M | 12000 | 50.4 | 00:00:00 + | | | | | | | | streeeeet2,+| | | | | + | | | | | | | | city2 | | | | | +(1 row) + +-- update first column +UPDATE update_test SET bser = bser - 1 + 1; +-- update middle column +UPDATE update_test SET perf_f = 8.9; +-- update last column +UPDATE update_test SET ctime = '00:00:00.1'; +-- update 3 continuos columns +UPDATE update_test SET destn = 'dcy2', samba = 0 WHERE Gend = 'M' and dept = 'CSD'; +-- update two non continuos columns +UPDATE update_test SET destn = 'moved', samba = 0; +UPDATE update_test SET bln = FALSE, hgt = 10.1; +-- update causing some column alignment difference +UPDATE update_test SET ename = 'Tes'; +UPDATE update_test SET dept = 'Test'; +SELECT * from update_test; + bser | bln | ename | perf_f | grade | dept | dob | idnum | addr | destn | gend | samba | hgt | ctime +------+-----+-------+--------+-------+-------+------------+-------+-----------------------------+--------+------+-------+------+------------ + 1 | f | Tes | 8.9 | B | Test | 01-01-2000 | 520 | road2, +| moved | M | 0 | 10.1 | 00:00:00.1 + | | | | | | | | streeeeet2,+| | | | | + | | | | | | | | city2 | | | | | +(1 row) + +DROP TABLE update_test; diff --git a/src/test/regress/sql/update.sql b/src/test/regress/sql/update.sql index a8a028f..1806992 100644 --- a/src/test/regress/sql/update.sql +++ b/src/test/regress/sql/update.sql @@ -59,3 +59,70 @@ UPDATE update_test SET c = repeat('x', 10000) WHERE c = 'car'; SELECT a, b, char_length(c) FROM update_test; DROP TABLE update_test; + + +-- +-- Test to update continuos and non continuos columns +-- + +DROP TABLE IF EXISTS update_test; +CREATE TABLE update_test ( + bser bigserial, + bln boolean, + ename VARCHAR(25), + perf_f float(8), + grade CHAR, + dept CHAR(5) NOT NULL, + dob DATE, + idnum INT, + addr VARCHAR(30) NOT NULL, + destn CHAR(6), + Gend CHAR, + samba BIGINT, + hgt float, + ctime TIME +); + +INSERT INTO update_test VALUES ( + nextval('update_test_bser_seq'::regclass), + TRUE, + 'Test', + 7.169, + 'B', + 'CSD', + '2000-01-01', + 520, + 'road2, + streeeeet2, + city2', + 'dcy2', + 'M', + 12000, + 50.4, + '00:00:00.0' +); + +SELECT * from update_test; + +-- update first column +UPDATE update_test SET bser = bser - 1 + 1; + +-- update middle column +UPDATE update_test SET perf_f = 8.9; + +-- update last column +UPDATE update_test SET ctime = '00:00:00.1'; + +-- update 3 continuos columns +UPDATE update_test SET destn = 'dcy2', samba = 0 WHERE Gend = 'M' and dept = 'CSD'; + +-- update two non continuos columns +UPDATE update_test SET destn = 'moved', samba = 0; +UPDATE update_test SET bln = FALSE, hgt = 10.1; + +-- update causing some column alignment difference +UPDATE update_test SET ename = 'Tes'; +UPDATE update_test SET dept = 'Test'; + +SELECT * from update_test; +DROP TABLE update_test;