/** Creates the {@link Scanner} to use for this query. */
Scanner getScanner() throws HBaseException {
final short metric_width = tsdb.metrics.width();
final byte[] start_row = new byte[metric_width + Const.TIMESTAMP_BYTES];
final byte[] end_row = new byte[metric_width + Const.TIMESTAMP_BYTES];
// We search at least one row before and one row after the start & end
// time we've been given as it's quite likely that the exact timestamp
// we're looking for is in the middle of a row. Plus, a number of things
// rely on having a few extra data points before & after the exact start
// & end dates in order to do proper rate calculation or downsampling near
// the "edges" of the graph.
Bytes.setInt(start_row, (int) getScanStartTime(), metric_width);
Bytes.setInt(
end_row,
(end_time == UNSET
? -1 // Will scan until the end (0xFFF...).
: (int) getScanEndTime()),
metric_width);
System.arraycopy(metric, 0, start_row, 0, metric_width);
System.arraycopy(metric, 0, end_row, 0, metric_width);
final Scanner scanner = tsdb.client.newScanner(tsdb.table);
scanner.setStartKey(start_row);
scanner.setStopKey(end_row);
if (tags.size() > 0 || group_bys != null) {
createAndSetFilter(scanner);
}
scanner.setFamily(TSDB.FAMILY);
return scanner;
}
/**
* Finds all the {@link Span}s that match this query. This is what actually scans the HBase table
* and loads the data into {@link Span}s.
*
* @return A map from HBase row key to the {@link Span} for that row key. Since a {@link Span}
* actually contains multiple HBase rows, the row key stored in the map has its timestamp
* zero'ed out.
* @throws HBaseException if there was a problem communicating with HBase to perform the search.
* @throws IllegalArgumentException if bad data was retreived from HBase.
*/
private TreeMap<byte[], Span> findSpans() throws HBaseException {
final short metric_width = tsdb.metrics.width();
final TreeMap<byte[], Span> spans = // The key is a row key from HBase.
new TreeMap<byte[], Span>(new SpanCmp(metric_width));
int nrows = 0;
int hbase_time = 0; // milliseconds.
long starttime = System.nanoTime();
final Scanner scanner = getScanner();
try {
ArrayList<ArrayList<KeyValue>> rows;
while ((rows = scanner.nextRows().joinUninterruptibly()) != null) {
hbase_time += (System.nanoTime() - starttime) / 1000000;
for (final ArrayList<KeyValue> row : rows) {
final byte[] key = row.get(0).key();
if (Bytes.memcmp(metric, key, 0, metric_width) != 0) {
throw new IllegalDataException(
"HBase returned a row that doesn't match"
+ " our scanner ("
+ scanner
+ ")! "
+ row
+ " does not start"
+ " with "
+ Arrays.toString(metric));
}
Span datapoints = spans.get(key);
if (datapoints == null) {
datapoints = new Span(tsdb);
spans.put(key, datapoints);
}
datapoints.addRow(tsdb.compact(row));
nrows++;
starttime = System.nanoTime();
}
}
} catch (RuntimeException e) {
throw e;
} catch (Exception e) {
throw new RuntimeException("Should never be here", e);
} finally {
hbase_time += (System.nanoTime() - starttime) / 1000000;
scanlatency.add(hbase_time);
}
LOG.info(this + " matched " + nrows + " rows in " + spans.size() + " spans");
if (nrows == 0) {
return null;
}
return spans;
}
/**
* Sets the server-side regexp filter on the scanner. In order to find the rows with the relevant
* tags, we use a server-side filter that matches a regular expression on the row key.
*
* @param scanner The scanner on which to add the filter.
*/
void createAndSetFilter(final Scanner scanner) {
if (group_bys != null) {
Collections.sort(group_bys, Bytes.MEMCMP);
}
final short name_width = tsdb.tag_names.width();
final short value_width = tsdb.tag_values.width();
final short tagsize = (short) (name_width + value_width);
// Generate a regexp for our tags. Say we have 2 tags: { 0 0 1 0 0 2 }
// and { 4 5 6 9 8 7 }, the regexp will be:
// "^.{7}(?:.{6})*\\Q\000\000\001\000\000\002\\E(?:.{6})*\\Q\004\005\006\011\010\007\\E(?:.{6})*$"
final StringBuilder buf =
new StringBuilder(
15 // "^.{N}" + "(?:.{M})*" + "$"
+ ((13 + tagsize) // "(?:.{M})*\\Q" + tagsize bytes + "\\E"
* (tags.size() + (group_bys == null ? 0 : group_bys.size() * 3))));
// In order to avoid re-allocations, reserve a bit more w/ groups ^^^
// Alright, let's build this regexp. From the beginning...
buf.append(
"(?s)" // Ensure we use the DOTALL flag.
+ "^.{")
// ... start by skipping the metric ID and timestamp.
.append(tsdb.metrics.width() + Const.TIMESTAMP_BYTES)
.append("}");
final Iterator<byte[]> tags = this.tags.iterator();
final Iterator<byte[]> group_bys =
(this.group_bys == null ? new ArrayList<byte[]>(0).iterator() : this.group_bys.iterator());
byte[] tag = tags.hasNext() ? tags.next() : null;
byte[] group_by = group_bys.hasNext() ? group_bys.next() : null;
// Tags and group_bys are already sorted. We need to put them in the
// regexp in order by ID, which means we just merge two sorted lists.
do {
// Skip any number of tags.
buf.append("(?:.{").append(tagsize).append("})*\\Q");
if (isTagNext(name_width, tag, group_by)) {
addId(buf, tag);
tag = tags.hasNext() ? tags.next() : null;
} else { // Add a group_by.
addId(buf, group_by);
final byte[][] value_ids = (group_by_values == null ? null : group_by_values.get(group_by));
if (value_ids == null) { // We don't want any specific ID...
buf.append(".{").append(value_width).append('}'); // Any value ID.
} else { // We want specific IDs. List them: /(AAA|BBB|CCC|..)/
buf.append("(?:");
for (final byte[] value_id : value_ids) {
buf.append("\\Q");
addId(buf, value_id);
buf.append('|');
}
// Replace the pipe of the last iteration.
buf.setCharAt(buf.length() - 1, ')');
}
group_by = group_bys.hasNext() ? group_bys.next() : null;
}
} while (tag != group_by); // Stop when they both become null.
// Skip any number of tags before the end.
buf.append("(?:.{").append(tagsize).append("})*$");
scanner.setKeyRegexp(buf.toString(), CHARSET);
}
private static int fsck(
final TSDB tsdb,
final HBaseClient client,
final byte[] table,
final boolean fix,
final String[] args)
throws Exception {
/** Callback to asynchronously delete a specific {@link KeyValue}. */
final class DeleteOutOfOrder implements Callback<Deferred<Object>, Object> {
private final KeyValue kv;
public DeleteOutOfOrder(final KeyValue kv) {
this.kv = kv;
}
public Deferred<Object> call(final Object arg) {
return client.delete(new DeleteRequest(table, kv.key(), kv.family(), kv.qualifier()));
}
public String toString() {
return "delete out-of-order data";
}
}
int errors = 0;
int correctable = 0;
final short metric_width = width(tsdb, "metrics");
final short name_width = width(tsdb, "tag_names");
final short value_width = width(tsdb, "tag_values");
final ArrayList<Query> queries = new ArrayList<Query>();
CliQuery.parseCommandLineQuery(args, tsdb, queries, null, null);
final StringBuilder buf = new StringBuilder();
for (final Query query : queries) {
final long start_time = System.nanoTime();
long ping_start_time = start_time;
LOG.info("Starting to fsck data covered by " + query);
long kvcount = 0;
long rowcount = 0;
final Bytes.ByteMap<Seen> seen = new Bytes.ByteMap<Seen>();
final Scanner scanner = Core.getScanner(query);
ArrayList<ArrayList<KeyValue>> rows;
while ((rows = scanner.nextRows().joinUninterruptibly()) != null) {
for (final ArrayList<KeyValue> row : rows) {
rowcount++;
// Take a copy of the row-key because we're going to zero-out the
// timestamp and use that as a key in our `seen' map.
final byte[] key = row.get(0).key().clone();
final long base_time = Bytes.getUnsignedInt(key, metric_width);
for (int i = metric_width; i < metric_width + Const.TIMESTAMP_BYTES; i++) {
key[i] = 0;
}
Seen prev = seen.get(key);
if (prev == null) {
prev = new Seen(base_time - 1, row.get(0));
seen.put(key, prev);
}
for (final KeyValue kv : row) {
kvcount++;
if (kvcount % 100000 == 0) {
final long now = System.nanoTime();
ping_start_time = (now - ping_start_time) / 1000000;
LOG.info(
"... "
+ kvcount
+ " KV analyzed in "
+ ping_start_time
+ "ms ("
+ (100000 * 1000 / ping_start_time)
+ " KVs/s)");
ping_start_time = now;
}
if (kv.qualifier().length != 2) {
LOG.warn(
"Ignoring unsupported KV with a qualifier of "
+ kv.qualifier().length
+ " bytes:"
+ kv);
continue;
}
final short qualifier = Bytes.getShort(kv.qualifier());
final short delta = (short) ((qualifier & 0xFFFF) >>> FLAG_BITS);
final long timestamp = base_time + delta;
byte[] value = kv.value();
if (value.length > 8) {
errors++;
LOG.error("Value more than 8 byte long with a 2-byte" + " qualifier.\n\t" + kv);
}
// TODO(tsuna): Don't hardcode 0x8 / 0x3 here.
if ((qualifier & (0x8 | 0x3)) == (0x8 | 0x3)) { // float | 4 bytes
// The qualifier says the value is on 4 bytes, and the value is
// on 8 bytes, then the 4 MSBs must be 0s. Old versions of the
// code were doing this. It's kinda sad. Some versions had a
// bug whereby the value would be sign-extended, so we can
// detect these values and fix them here.
if (value.length == 8) {
if (value[0] == -1 && value[1] == -1 && value[2] == -1 && value[3] == -1) {
errors++;
correctable++;
if (fix) {
value = value.clone(); // We're going to change it.
value[0] = value[1] = value[2] = value[3] = 0;
client.put(new PutRequest(table, kv.key(), kv.family(), kv.qualifier(), value));
} else {
LOG.error(
"Floating point value with 0xFF most significant"
+ " bytes, probably caused by sign extension bug"
+ " present in revisions [96908436..607256fc].\n"
+ "\t"
+ kv);
}
} else if (value[0] != 0 || value[1] != 0 || value[2] != 0 || value[3] != 0) {
errors++;
}
} else if (value.length != 4) {
errors++;
LOG.error(
"This floating point value must be encoded either on"
+ " 4 or 8 bytes, but it's on "
+ value.length
+ " bytes.\n\t"
+ kv);
}
}
if (timestamp <= prev.timestamp()) {
errors++;
correctable++;
if (fix) {
final byte[] newkey = kv.key().clone();
// Fix the timestamp in the row key.
final long new_base_time = (timestamp - (timestamp % Const.MAX_TIMESPAN));
Bytes.setInt(newkey, (int) new_base_time, metric_width);
final short newqual =
(short) ((timestamp - new_base_time) << FLAG_BITS | (qualifier & FLAGS_MASK));
final DeleteOutOfOrder delooo = new DeleteOutOfOrder(kv);
if (timestamp < prev.timestamp()) {
client
.put(
new PutRequest(
table, newkey, kv.family(), Bytes.fromShort(newqual), value))
// Only delete the offending KV once we're sure that the new
// KV has been persisted in HBase.
.addCallbackDeferring(delooo);
} else {
// We have two data points at exactly the same timestamp.
// This can happen when only the flags differ. This is
// typically caused by one data point being an integer and
// the other being a floating point value. In this case
// we just delete the duplicate data point and keep the
// first one we saw.
delooo.call(null);
}
} else {
buf.setLength(0);
buf.append(
timestamp < prev.timestamp()
? "Out of order data.\n\t"
: "Duplicate data point with different flags.\n\t")
.append(timestamp)
.append(" (")
.append(DumpSeries.date(timestamp))
.append(") @ ")
.append(kv)
.append("\n\t");
DumpSeries.formatKeyValue(buf, tsdb, kv, base_time);
buf.append("\n\t was found after\n\t")
.append(prev.timestamp)
.append(" (")
.append(DumpSeries.date(prev.timestamp))
.append(") @ ")
.append(prev.kv)
.append("\n\t");
DumpSeries.formatKeyValue(
buf, tsdb, prev.kv, Bytes.getUnsignedInt(prev.kv.key(), metric_width));
LOG.error(buf.toString());
}
} else {
prev.setTimestamp(timestamp);
prev.kv = kv;
}
}
}
}
final long timing = (System.nanoTime() - start_time) / 1000000;
System.out.println(
kvcount
+ " KVs (in "
+ rowcount
+ " rows) analyzed in "
+ timing
+ "ms (~"
+ (kvcount * 1000 / timing)
+ " KV/s)");
}
System.out.println(errors != 0 ? "Found " + errors + " errors." : "No error found.");
if (!fix && correctable > 0) {
System.out.println(
correctable
+ " of these errors are automatically"
+ " correctable, re-run with --fix.\n"
+ "Make sure you understand the errors above and you"
+ " know what you're doing before using --fix.");
}
return errors;
}
