diff --git a/api.py b/api.py
new file mode 100644
index 0000000..f2efdbb
--- /dev/null
+++ b/api.py
@@ -0,0 +1,163 @@
+"""HTTP API implementation."""
+
+import csv
+import json
+import logging
+import time
+import webapp2
+
+import models
+
+
+def Multiplex(iterators, selection_func):
+  """Iterate over multiple sources and multiplex values.
+
+  Args:
+    iterators: list of source iterators
+    selection_func: takes a list of values, selects one and returns its index
+  Yields:
+    Each value from the source iterators, in the order determined by
+    selection_func.
+  """
+  iterators = list(iterators)
+
+  for i, obj in reversed(list(enumerate(iterators))):
+    try:
+      iterators[i] = [obj, obj.next()]
+    except StopIteration:
+      del iterators[i]
+
+  while iterators:
+    i = selection_func([x[1] for x in iterators])
+    yield iterators[i][1]
+    try:
+      iterators[i][1] = iterators[i][0].next()
+    except StopIteration:
+      del iterators[i]
+
+
+class Get(webapp2.RequestHandler):
+  """Fetch values from one or more timeseries.
+
+  HTTP parameters:
+    expr=<string>
+    format={csv,json}
+    resolution={full,minute,hour,day}
+    start=<unix_timestamp>
+    end=<unix_timestamp>
+  """
+
+  def get(self):
+    output_format = self.request.get('format', 'json')
+    assert output_format in ('csv', 'json')
+
+    resolution = self.request.get('resolution', 'full')
+    resolution = models.Values.RESOLUTION_NAMES[resolution]
+
+    start = self.request.get('start', None)
+    if start:
+      start = int(start)
+      if start < 0:
+        start = int(time.time()) + start
+    end = self.request.get('end', None)
+    if end:
+      end = int(end)
+      if end < 0:
+        end = int(time.time()) + end
+
+    expr = self.request.get('expr')
+    data = models.TimeSeries.FromExpr(expr)
+
+    if output_format == 'csv':
+      self.response.content_type = 'text/csv'
+      fh = csv.writer(self.response.out)
+      # CSV requires us to pre-determine column names
+      try:
+        group = data.iterkeys().next()
+      except StopIteration:
+        return
+      group_columns = [x.tag_key for x in group]
+      fh.writerow(['timestamp'] + group_columns + ['value'])
+    elif output_format == 'json':
+      self.response.content_type = 'application/json'
+      ret = []
+
+    def SelectMinTime(values):
+      return min(range(len(values)), key=lambda x: values[x][0])
+
+    streams = {}
+
+    # Get all datastore queries running in parallel first
+    for groupings, timeseries in data.iteritems():
+      streams[groupings] = [
+          x.GetValues(start=start, end=end, resolution=resolution)
+          for x in timeseries
+      ]
+
+    for groupings in sorted(data.keys()):
+      group_values = dict((x.tag_key, x.tag_value) for x in sorted(groupings))
+
+      values = Multiplex(streams[groupings], SelectMinTime)
+
+      if output_format == 'csv':
+        group_constants = [group_values[x] for x in group_columns]
+        for timestamp, value in values:
+          fh.writerow([timestamp] + group_constants + [value])
+      elif output_format == 'json':
+        ret.append({
+            'tags': group_values,
+            'timestamps_values': list(values),
+        })
+
+    if output_format == 'json':
+      json.dump(ret, self.response.out, separators=(',', ':'))
+
+
+class Put(webapp2.RequestHandler):
+  """Add a value to a timeseries.
+
+  HTTP parameters:
+    tag=<key>=<value> (repeated)
+    value=<integer>
+  """
+
+  def _HandleBlock(self, block):
+    if 'timestamps_values' not in block:
+      now = int(time.time())
+      block['timestamps_values'] = [[now, x] for x in block['values']]
+
+    if 'client_timestamp' in block:
+      server_timestamp = int(time.time())
+      offset = server_timestamp - block['client_timestamp']
+      for pair in block['timestamps_values']:
+        pair[0] += offset
+
+    tags = [models.Tag.FromStr(x, create=True)
+            for x in block['tags']]
+    timeseries = models.TimeSeries.GetOrCreate(tags)
+
+    timeseries.AddValues(block['timestamps_values'],
+                         offset=block.get('offset', False))
+
+  def post(self):
+    content_type = self.request.headers['Content-Type'].split(';', 1)[0]
+
+    if content_type == 'application/x-www-form-urlencoded':
+      self._HandleBlock({
+          'offset': bool(self.request.get('offset', 0)),
+          'tags': self.request.get_all('tag'),
+          'values': [int(x) for x in self.request.get_all('value')],
+      })
+
+    elif content_type == 'application/json':
+      for block in json.loads(self.request.body):
+        self._HandleBlock(block)
+
+    else:
+      assert False, content_type
+
+
+app = webapp2.WSGIApplication([
+    ('/api/get', Get),
+    ('/api/put', Put),
+])
diff --git a/app.yaml b/app.yaml
new file mode 100644
index 0000000..76d8669
--- /dev/null
+++ b/app.yaml
@@ -0,0 +1,20 @@
+application: yatsdb
+version: 1
+runtime: python27
+api_version: 1
+threadsafe: true
+
+handlers:
+- url: /api/.*
+  script: api.app
+  secure: always
+
+- url: /console
+  static_files: console/console.html
+  upload: console/console.html
+
+- url: /console
+  static_dir: console
+
+builtins:
+- remote_api: on
diff --git a/console/console.css b/console/console.css
new file mode 100644
index 0000000..6aa168d
--- /dev/null
+++ b/console/console.css
@@ -0,0 +1,41 @@
+@import url(https://fonts.googleapis.com/css?family=Roboto+Mono);
+
+body {
+  margin: 0;
+}
+
+.expr {
+  position: fixed;
+  top: 0;
+  left: 0;
+  right: 0;
+  z-index: 1;
+}
+
+.expr form {
+  margin: 0;
+}
+
+.expr input {
+  width: 100%;
+  border: 0;
+  padding: 5px;
+  font-family: 'Roboto Mono';
+  font-size: 15px;
+  border-bottom: 1px solid #757575;
+  background: rgba(255,255,255,0.7);
+}
+
+.expr input:focus {
+  outline: none;
+  border-bottom: 1px solid #5264AE;
+}
+
+.chart {
+  width: 900px;
+  height: 500px;
+  display: inline-block;
+  -moz-user-select: none;
+  -webkit-user-select: none;
+  -ms-user-select: none;
+}
diff --git a/console/console.html b/console/console.html
new file mode 100644
index 0000000..de1207b
--- /dev/null
+++ b/console/console.html
@@ -0,0 +1,12 @@
+<html>
+  <head>
+    <title>YATSDB Console</title>
+    <script type="text/javascript" src="https://www.google.com/jsapi"></script>
+    <script src="/console/lib.js"></script>
+    <script src="/console/console.js"></script>
+    <link rel="stylesheet" type="text/css" href="/console/console.css">
+  </head>
+  <body>
+    <div id="container"></div>
+  </body>
+</html>
diff --git a/console/console.js b/console/console.js
new file mode 100644
index 0000000..4ea249e
--- /dev/null
+++ b/console/console.js
@@ -0,0 +1,175 @@
+var Console = function(container) {
+  this.container_ = container;
+  this.client_ = new TSDBClient();
+  this.loadJSAPI_();
+};
+
+
+Console.prototype.loadJSAPI_ = function() {
+  google.load('visualization', '1.1', {
+    packages: ['corechart'],
+    callback: this.jsapiLoaded_.bind(this),
+  });
+};
+
+
+Console.prototype.jsapiLoaded_ = function() {
+  var exprContainer = document.createElement('div');
+  exprContainer.className = 'expr';
+  this.container_.appendChild(exprContainer);
+
+  var form = document.createElement('form');
+  exprContainer.appendChild(form);
+
+  this.exprInput_ = document.createElement('input');
+  form.appendChild(this.exprInput_);
+  form.addEventListener('submit', this.onExprChange_.bind(this));
+
+  var chartContainer = document.createElement('div');
+  this.container_.appendChild(chartContainer);
+
+  this.charts_ = [
+    {
+      'title': 'Last 20 minutes',
+      'seconds': 120 * 10,
+      'resolution': 'full',
+    },
+    {
+      'title': 'Last 2 hours',
+      'seconds': 120 * 60,
+      'resolution': 'minute',
+    },
+    {
+      'title': 'Last 5 days',
+      'seconds': 120 * 60 * 60,
+      'resolution': 'hour',
+    },
+    {
+      'title': 'Last 120 days',
+      'seconds': 120 * 60 * 60 * 24,
+      'resolution': 'day',
+    },
+  ];
+  for (var i = 0; i < this.charts_.length; i++) {
+    var obj = this.charts_[i];
+    obj.container = document.createElement('div');
+    obj.container.className = 'chart';
+    chartContainer.appendChild(obj.container);
+  }
+
+  this.expr_ = decodeURIComponent(window.location.hash.substring(1));
+  // Ensure that the URL is properly encoded for copy & paste.
+  window.location.hash = encodeURIComponent(this.expr_);
+  this.exprInput_.value = this.expr_;
+  window.addEventListener('hashchange', this.onHashChange_.bind(this));
+
+  this.loadCharts_();
+};
+
+
+Console.prototype.onExprChange_ = function(e) {
+  window.location.hash = encodeURIComponent(this.exprInput_.value);
+  e.preventDefault();
+};
+
+
+Console.prototype.onHashChange_ = function(e) {
+  this.expr_ = decodeURIComponent(window.location.hash.substring(1));
+  this.exprInput_.value = this.expr_;
+  this.loadCharts_();
+};
+
+
+Console.prototype.loadCharts_ = function() {
+  if (!this.expr_) {
+    return;
+  }
+
+  for (var i = 0; i < this.charts_.length; i++) {
+    var options = this.charts_[i];
+    if (options.instance) {
+      options.instance.destroy();
+    }
+    options.instance = new Chart(this.client_, this.expr_, options);
+  }
+};
+
+
+var Chart = function(client, expr, options) {
+  this.options_ = options;
+  this.watch_ = client.watch(
+      expr, this.options_.seconds, this.options_.resolution,
+      this.drawChart_.bind(this));
+};
+
+
+Chart.prototype.drawChart_ = function(data) {
+  var dataTable = new google.visualization.DataTable();
+  dataTable.addColumn('date', 'Timestamp');
+
+  for (var i = data.length - 1; i >= 0; i--) {
+    var ts = data[i];
+    if (!ts.timestamps_values.length) {
+      data.splice(i, 1);
+      continue;
+    }
+    var tags = [];
+    for (var key in ts.tags) {
+      tags.push(key + '=' + ts.tags[key]);
+    }
+    dataTable.addColumn('number', tags.join(','));
+  }
+  for (var i = 0; i < data.length; i++) {
+    var ts = data[i];
+    for (var j = 0; j < ts.timestamps_values.length; j++) {
+      var row = [new Date(ts.timestamps_values[j][0] * 1000)];
+      for (var k = 0; k < i; k++) {
+        row.push(null);
+      }
+      row.push(ts.timestamps_values[j][1]);
+      for (var k = i + 1; k < data.length; k++) {
+        row.push(null);
+      }
+      dataTable.addRow(row);
+    }
+  };
+
+  var options = {
+    title: this.options_.title,
+    legend: {
+      position: 'bottom'
+    },
+    hAxis: {
+      gridlines: {
+        count: -1,
+      },
+    },
+    explorer: {
+      actions: [
+          'dragToZoom',
+          'rightClickToReset',
+      ],
+    },
+  };
+
+  // It'd be great to use the material design charts here, but they fail to load
+  // ~25% of the time, which is a non-starter.
+  if (this.chartObj_) {
+    this.chartObj_.clearChart();
+  }
+  this.chartObj_ = new google.visualization.LineChart(this.options_.container);
+  this.chartObj_.draw(dataTable, options);
+};
+
+
+Chart.prototype.destroy = function() {
+  this.watch_.destroy();
+  if (this.chartObj_) {
+    this.chartObj_.clearChart();
+  }
+};
+
+
+document.addEventListener('DOMContentLoaded', function(e) {
+  new Console(document.getElementById('container'));
+});
diff --git a/console/lib.js b/console/lib.js
new file mode 100644
index 0000000..14fd7b2
--- /dev/null
+++ b/console/lib.js
@@ -0,0 +1,78 @@
+var TSDBClient = function(opt_baseURL) {
+  this.baseURL_ = opt_baseURL || '/api';
+};
+
+
+TSDBClient.prototype.watch = function(
+    expr, windowSeconds, resolution, callback) {
+  return new TSDBWatch(
+      this.baseURL_, expr, windowSeconds, resolution, callback);
+};
+
+
+var TSDBWatch = function(baseURL, expr, windowSeconds, resolution, callback) {
+  this.url_ = baseURL + '/get?start=-' + windowSeconds.toString() +
+      '&resolution=' + encodeURIComponent(resolution) +
+      '&expr=' + encodeURIComponent(expr);
+  this.callback_ = callback;
+  this.refresh_ = windowSeconds / 40;
+  this.delay_ = TSDBWatch.MIN_DELAY;
+  this.sendRequest_();
+};
+
+
+TSDBWatch.MIN_DELAY = 0.5;
+TSDBWatch.MAX_DELAY = 32.0;
+TSDBWatch.DELAY_MULT = 2;
+
+
+TSDBWatch.prototype.destroy = function() {
+  if (this.timer_) {
+    window.clearTimeout(this.timer_);
+  }
+};
+
+
+TSDBWatch.prototype.sendRequest_ = function() {
+  this.timer_ = null;
+
+  var xhr = new XMLHttpRequest();
+  xhr.responseType = 'json';
+  xhr.timeout = 30 * 1000;
+  xhr.open('GET', this.url_);
+  xhr.addEventListener('load', this.onLoad_.bind(this));
+  xhr.addEventListener('error', this.retry_.bind(this));
+  xhr.addEventListener('timeout', this.retry_.bind(this));
+  xhr.send();
+};
+
+
+TSDBWatch.prototype.retry_ = function(e) {
+  if (this.timer_) {
+    console.log('Duplicate retry call');
+    return;
+  }
+  this.timer_ = window.setTimeout(
+      this.sendRequest_.bind(this),
+      this.delay_ * 1000);
+  this.delay_ = Math.min(this.delay_ * TSDBWatch.DELAY_MULT,
+                         TSDBWatch.MAX_DELAY);
+};
+
+
+TSDBWatch.prototype.onLoad_ = function(e) {
+  var data = e.target.response;
+
+  if (!data) {
+    this.retry_();
+    return;
+  }
+
+  this.delay_ = TSDBWatch.MIN_DELAY;
+
+  this.callback_(data);
+
+  this.timer_ = window.setTimeout(
+      this.sendRequest_.bind(this),
+      this.refresh_ * 1000);
+}
diff --git a/models.py b/models.py
new file mode 100644
index 0000000..c70eac0
--- /dev/null
+++ b/models.py
@@ -0,0 +1,761 @@
+"""Datastore model definitions and utility functions."""
+
+import bisect
+import collections
+import itertools
+import logging
+import re
+import StringIO
+import sys
+import time
+
+from google.appengine.api import memcache
+from google.appengine.ext import db
+
+
+class Error(Exception):
+  pass
+
+
+class InvalidExpression(Error):
+  pass
+
+
+class InvalidSerializedData(Error):
+  pass
+
+
+class Tag(db.Model):
+  """A single unique key/value tag pair.
+
+  Immutable once written, so positively cacheable.
+  """
+  # key_name = string representation of tag key and value
+
+  tag_key = db.StringProperty(required=True)
+  tag_value = db.StringProperty(required=True)
+
+  _cache = {}
+
+  @classmethod
+  def Get(cls, key, value):
+    """Fetch a single Tag.
+
+    May fetch from cache or datastore. Positive but not negative caching.
+
+    Args:
+      key: string
+      value: string
+    Returns:
+      Tag object, or None.
+    """
+    cache_key = '%s=%s' % (key, value)
+    obj = cls._cache.get(cache_key, None)
+    if obj:
+      return obj
+    obj = cls.get_by_key_name(cache_key)
+    if obj:
+      cls._cache[cache_key] = obj
+      return obj
+    return None
+
+  @classmethod
+  def GetOrCreate(cls, key, value):
+    """Fetch or create a single Tag.
+
+    May fetch from cache or datastore. Positive but not negative caching.
+
+    Args:
+      key: string
+      value: string
+    Returns:
+      Tag object.
+    """
+    cache_key = '%s=%s' % (key, value)
+    obj = cls._cache.get(cache_key, None)
+    if obj:
+      return obj
+    obj = cls.get_or_insert(cache_key, tag_key=key, tag_value=value)
+    cls._cache[cache_key] = obj
+    return obj
+
+  @classmethod
+  def FromStr(cls, tagstring, create=False):
+    """Fetch a Tag using a key=value string representation.
+
+    Args:
+      tagstring: <key>=<value>
+      create: if True, create the Tag if it doesn't exist. Only use from POST
+        requests.
+    Returns:
+      Tag object, or None (if create=False).
+    """
+    key, value = tagstring.split('=', 1)
+    if create:
+      return cls.GetOrCreate(key, value)
+    else:
+      return cls.Get(key, value)
+
+  @classmethod
+  def _CacheAndYield(cls, query):
+    """Fetch and yield tags from a query, caching the results sa we go.
+
+    Args:
+      query: already-running datastore query that returns Tag objects
+    Yields:
+      Tag objects
+    """
+    for tag in query:
+      cls._cache[tag.key().name()] = tag
+      yield tag
+
+  @classmethod
+  def FromKey(cls, keystring):
+    """Fetch all tags with a given key.
+
+    Args:
+      keystring: string
+    Returns:
+      Iterator of Tag objects.
+    """
+    query = (cls.all()
+             .filter('tag_key =', keystring)
+             .run())
+    return cls._CacheAndYield(query)
+
+  def __cmp__(self, other):
+    return cmp((self.tag_key, self.tag_value),
+               (other.tag_key, other.tag_value))
+
+  def __str__(self):
+    return '%s=%s' % (self.tag_key, self.tag_value)
+
+  def __repr__(self):
+    return str(self)
+
+
+class TimeSeries(db.Model):
+  """A unique set of tags.
+
+  Immutable once written, so positively cacheable.
+  """
+  _EXPR_RE = re.compile('^\\{(?P<selectors>.*)\\}(\\[(?P<groupings>.*)\\])?$')
+
+  # key_name = string representation of sorted tags
+  tags = db.ListProperty(item_type=db.Key, required=True)
+
+  _cache = {}
+
+  @classmethod
+  def GetOrCreate(cls, tags):
+    """Fetch a single TimeSeries.
+
+    May fetch from cache or datastore. Positive but not negative caching.
+
+    Args:
+      tags: list of Tag objects
+    Returns:
+      TimeSeries object.
+    """
+    tags.sort(key=lambda x: (x.tag_key, x.tag_value))
+    cache_key = ','.join(str(x) for x in tags)
+    obj = cls._cache.get(cache_key, None)
+    if obj:
+      return obj
+    obj = cls.get_or_insert(cache_key, tags=[x.key() for x in tags])
+    cls._cache[cache_key] = obj
+    return obj
+
+  @classmethod
+  def FromExpr(cls, expr):
+    """Fetch TimeSeries using a string expression.
+
+    Expression format is:
+      {<key1>=<value1>,<key2>=<value2>,...}[<key3>,<key4>,...]
+
+    Args:
+      expr: string, see format above
+    Returns:
+      dict of (grouping key,grouping key,...): list of TimeSeries objects
+      Empty tuple key if there are no grouping keys
+    Raises:
+      InvalidExpression: expr not parseable
+    """
+    parsed = cls._EXPR_RE.match(expr)
+    if not parsed:
+      raise InvalidExpression
+
+    tags = []
+    for key_value in parsed.group('selectors').split(','):
+      tags.append(Tag.FromStr(key_value))
+
+    groupings = []
+    groupings_str = parsed.group('groupings')
+    if groupings_str:
+      for grouping in groupings_str.split(','):
+        groupings.append(Tag.FromKey(grouping))
+
+    return cls.GetPartial(tags, groupings)
+
+  @classmethod
+  def GetPartial(cls, tags, groupings):
+    """Get a list of TimeSeries matches for the given tags.
+
+    Does partial matching, e.g. key1=value1 matches the TimeSeries
+    key1=value1,key2=value2. This enables forwards-compatibility, i.e. it
+    allows data sources to add new tags without breaking existing readers.
+
+    Args:
+      tags: list of Tag objects, for selection
+      groupings: list of list of Tag objects, for grouping
+    Returns:
+      dict of (grouping key,grouping key,...): list of TimeSeries objects
+      Empty tuple key if there are no grouping keys
+    """
+    # Build all our query objects, then go through and use them, trying to
+    # keep a minimum set in memory at all times.
+
+    # Build and start queries for selectors.
+    tag_queries = []
+    for tag in tags:
+      tag_queries.append(cls.all(keys_only=True)
+                         .filter('tags =', tag)
+                         .run())
+
+    # Build and start queries for grouping. Also build a dict of sets used later
+    # to structure the response.
+    tags_by_tag_keys = collections.defaultdict(set)
+    grouping_queries = []
+    for grouping in groupings:
+      or_queries = []
+      for tag in grouping:
+        tags_by_tag_keys[tag.tag_key].add(tag.key())
+        or_queries.append(cls.all(keys_only=True)
+                          .filter('tags =', tag)
+                          .run())
+      grouping_queries.append(or_queries)
+
+    # Pull the results from the selection queries and combine them (AND).
+    keys = None
+    for query in tag_queries:
+      timeseries = set(query)
+      if keys is None:
+        keys = timeseries
+      else:
+        keys &= timeseries
+
+    # Pull the results from the grouping queries and combine them (OR within
+    # the tags for a group, AND between groups). We eliminate any TimeSeries
+    # that lack any values for the grouping tags.
+    for or_queries in grouping_queries:
+      or_set = set()
+      for query in or_queries:
+        timeseries = set(query)
+        # Pre-filter to trim down things to keep.
+        timeseries &= keys
+        or_set |= timeseries
+      keys &= or_set
+
+    # Fetch all the relevant TimeSeries objects, either from cache or in from
+    # datastore in batch. Write anything fetched from datastore back to cache.
+    timeseries = []
+    to_fetch = []
+    for key in keys:
+      obj = cls._cache.get(key.name(), None)
+      if obj:
+        timeseries.append(obj)
+      else:
+        to_fetch.append(key)
+    if to_fetch:
+      for obj in cls.get(to_fetch):
+        cls._cache[obj.key().name()] = obj
+        timeseries.append(obj)
+
+    # Shortcut for queries without grouping.
+    if not groupings:
+      return {
+          (): timeseries,
+      }
+
+    # Bucket the TimeSeries objects by their grouping values.
+    ret = collections.defaultdict(list)
+    for ts in timeseries:
+      ts_tags = set(ts.tags)
+      intersects = [
+          ts_tags & tags
+          for tags in tags_by_tag_keys.itervalues()
+      ]
+      # We handle cases where a given TimeSeries has multiple values for the
+      # same tag key. Those go into multiple buckets, hence product().
+      for keysets in itertools.product(*intersects):
+        tags = tuple(Tag.FromStr(x.name()) for x in keysets)
+        ret[tags].append(ts)
+
+    return ret
+
+  @classmethod
+  def KeyName(cls, obj):
+    if isinstance(obj, db.Model):
+      return obj.key().name()
+    elif isinstance(obj, db.Key):
+      return obj.name()
+    else:
+      assert False, obj
+
+  def AddValue(self, value, timestamp=None, offset=False):
+    """Add a value to this TimeSeries.
+
+    Finds or creates the appropriate Values child object and adds the new
+    value to it.
+
+    Args:
+      value: integer
+      timestamp: UNIX timestamp; defaults to now
+      offset: if True, values are offsets from previous value
+    """
+    timestamp = timestamp or int(time.time())
+    Values.AddValue(self, timestamp, value, offset=offset)
+
+  def AddValues(self, timestamp_value_pairs, offset=False):
+    """Add values to this TimeSeries.
+
+    Args:
+      timestamp_value_pairs: list of (unix_timestamp, value) tuples
+      offset: if True, values are offsets from previous value
+    """
+    if not timestamp_value_pairs:
+      return
+    Values.AddValues(self, timestamp_value_pairs, offset=offset)
+
+  def _FilterValues(self, values_list, start, end):
+    """Filter values based on criteria.
+
+    Separated to move the yield call out of GetValues().
+
+    See GetValues() for args.
+    """
+    for values in values_list:
+      for timestamp, value in values.GetValues():
+        if timestamp < start or timestamp > end:
+          continue
+        yield (timestamp, value)
+
+  def GetValues(self, start=None, end=None, resolution=None):
+    """Fetch values from this TimeSeries.
+
+    Args:
+      start: UNIX timestamp; defaults to 0
+      end: UNIX timestamp; defaults to sys.maxint
+      resolution: one of Values.RESOLUTIONS
+    Yields:
+      (unix_timestamp, value) pairs, in ascending chronological order
+    """
+    start = start or 0
+    end = end or sys.maxint
+    resolution = resolution or Values.FULL
+
+    memiter = []
+    if resolution in Values.LEADING_BLOCK_IN_MEMCACHE:
+      client = memcache.Client()
+      namespace = 'TimeSeries:%d' % resolution
+      values = client.get(self.KeyName(self), namespace=namespace)
+      if values:
+        if values.start_time <= start:
+          return self._FilterValues([values], start, end)
+        memiter = [values]
+
+    query1 = (Values.all()
+              .ancestor(self)
+              .filter('resolution =', resolution)
+              .filter('start_time <=', start)
+              .order('-start_time'))
+    values1 = query1.run(limit=1)
+
+    query2 = (Values.all()
+              .ancestor(self)
+              .filter('resolution =', resolution)
+              .filter('start_time >', start)
+              .filter('start_time <=', end)
+              .order('start_time'))
+    values2 = query2.run()
+
+    return self._FilterValues(itertools.chain(values1, values2, memiter),
+                              start, end)
+
+  def __str__(self):
+    return self.key().name()
+
+
+class Values(db.Model):
+  """A chunk of values for a timeseries.
+
+  Subject to read/modify/write which must be transactional.
+  """
+  # parent = TimeSeries
+
+  FULL = 0
+  MINUTE = 1
+  HOUR = 2
+  DAY = 3
+
+  SECONDS = {
+      MINUTE: 60,
+      HOUR: 60 * 60,
+      DAY: 60 * 60 * 24,
+  }
+
+  RESOLUTION_NAMES = {
+      'full': FULL,
+      'minute': MINUTE,
+      'hour': HOUR,
+      'day': DAY,
+  }
+
+  RESOLUTIONS = {FULL, MINUTE, HOUR, DAY}
+  DOWNSAMPLES = {MINUTE, HOUR, DAY}
+
+  # Dict of resolution -> (max seconds in memcache, min seconds in memcache)
+  LEADING_BLOCK_IN_MEMCACHE = {
+      FULL: (60 * 60 * 24, 20 * 60),
+      MINUTE: (60 * 60 * 24, 2 * 60 * 60),
+  }
+
+  resolution = db.IntegerProperty(required=True, choices=RESOLUTIONS)
+
+  start_time = db.IntegerProperty(required=True)
+  start_value = db.IntegerProperty(required=True)
+
+  end_time = db.IntegerProperty(required=True)
+  end_value = db.IntegerProperty(required=True)
+
+  # Encoding: pairs of zigzag base128 varint encodings of deltas from the
+  # previous time (in seconds) and value (units unspecified)
+  times_and_values = db.BlobProperty()
+
+  _BLOB_LIMIT = 2**16
+
+  @classmethod
+  def AddValue(cls, timeseries, timestamp, value, offset=False,
+               resolution=FULL):
+    """Single value wrapper for AddValues()"""
+    return cls.AddValues(timeseries, [[timestamp, value]], offset, resolution)
+
+  @classmethod
+  def AddValues(cls, timeseries, timestamp_value_pairs, offset=False,
+                resolution=FULL):
+    """Find or create a block, then add values.
+
+    Args:
+      timeseries: parent TimeSeries object
+      timestamp_value_pairs: list of (unix_timestamp, value) tuples
+      offset: if True, values are offsets from previous value
+      resolution: one of Values.RESOLUTIONS
+    """
+    if not timestamp_value_pairs:
+      return
+    if resolution in cls.LEADING_BLOCK_IN_MEMCACHE:
+      cls._AddValuesMemcache(timeseries, timestamp_value_pairs, offset,
+                             resolution)
+    else:
+      cls._AddValuesDatastore(timeseries, timestamp_value_pairs, offset,
+                              resolution)
+
+  @classmethod
+  def _AddValuesMemcache(cls, timeseries, timestamp_value_pairs, offset,
+                         resolution):
+    """AddValues helper for memcache-backed blocks.
+
+    See AddValues() for arguments.
+    """
+    client = memcache.Client()
+    key = TimeSeries.KeyName(timeseries)
+    namespace = 'TimeSeries:%d' % resolution
+    values = client.gets(key, namespace=namespace)
+    if values:
+      values._AddValues(timestamp_value_pairs, offset=offset)
+      assert client.cas(key, values, namespace=namespace)
+    else:
+      values = cls._Create(timeseries, timestamp_value_pairs, resolution)
+      client.set(key, values, namespace=namespace)
+
+    max_age, min_age = cls.LEADING_BLOCK_IN_MEMCACHE[resolution]
+    now = int(time.time())
+    age = now - values.start_time
+    if age > max_age:
+      new_start = now - min_age
+      logging.info('Memcache block is too old (%d > %d); splitting at %d',
+                   age, max_age, new_start)
+      values._Split(split_timestamp=new_start)
+      client.set(key, values, namespace=namespace)
+
+  @classmethod
+  @db.transactional()
+  def _AddValuesDatastore(cls, timeseries, timestamp_value_pairs, offset,
+                          resolution, add_downsamples=True):
+    """AddValues helper for datastore-backed blocks.
+
+    See AddValues() for arguments.
+    """
+    values = (cls.all()
+              .ancestor(timeseries)
+              .filter('resolution =', resolution)
+              .order('-start_time')).fetch(1)
+    if values:
+      values = values[0]
+      values._AddValues(timestamp_value_pairs, offset=offset,
+                        add_downsamples=add_downsamples)
+    else:
+      values = cls._Create(timeseries, timestamp_value_pairs,
+                           resolution, add_downsamples=add_downsamples)
+
+    values.save()
+
+  @classmethod
+  def _Create(cls, timeseries, timestamp_value_pairs, resolution,
+              add_downsamples=True):
+    """Factory for new Values object."""
+    first_timestamp, first_value = timestamp_value_pairs[0]
+    values = Values(
+        parent=timeseries,
+        resolution=resolution,
+        start_time=first_timestamp,
+        start_value=first_value,
+        end_time=first_timestamp,
+        end_value=first_value)
+
+    if add_downsamples and resolution == cls.FULL:
+      # Add to downsamples just in case.
+      for downsample in cls.DOWNSAMPLES:
+        cls.AddValues(timeseries, [[first_timestamp, first_value]],
+                      resolution=downsample)
+
+    remaining = timestamp_value_pairs[1:]
+    values._AddValues(remaining, add_downsamples=add_downsamples)
+
+    return values
+
+  @classmethod
+  def ToZigZag(cls, value):
+    """Converts a value to zig zag encoding.
+
+    Args:
+      value: positive or negative integer
+    Returns:
+      Positive integer encoding
+    """
+    if value < 0:
+      return (abs(value) << 1) - 1
+    else:
+      return value << 1
+
+  @classmethod
+  def FromZigZag(cls, value):
+    """Converts a value from zig zag encoding.
+
+    Args:
+      value: positive encoded integer
+    Returns:
+      Positive or negative decoded integer
+    """
+    if value & 0x01:
+      return 0 - ((value + 1) >> 1)
+    else:
+      return value >> 1
+
+  @classmethod
+  def ToVarint(cls, value):
+    """Encodes an integer as a zig zag, variable-length string.
+
+    Args:
+      value: positive or negative integer
+    Returns:
+      A string with one or more encoded bytes
+    """
+    value = cls.ToZigZag(value)
+    ret = []
+    while value >= 2**7:
+      ret.append(chr(0x80 | (value % 2**7)))
+      value >>= 7
+    ret.append(chr(value))
+    return ''.join(ret)
+
+  @classmethod
+  def FromVarint(cls, fh):
+    """Decodes a variable-length, zig zag integer from a file handle.
+
+    Moves the file pointer forward to the byte after the encoded integer.
+
+    Args:
+      fh: file handle to read from; use StringIO to use a string
+    Returns:
+      Positive or negative decoded integer, or None if the stream stops
+        immediately
+    Raises:
+      InvalidSerializedData: stream stops in the middle of a varint
+    """
+    val = 0
+    shift = 0
+    while True:
+      byte = fh.read(1)
+      if not byte:
+        if val == 0:
+          return None
+        else:
+          raise InvalidSerializedData('incomplete varint')
+      byte_val = ord(byte)
+      val |= ((byte_val & 0x7f) << shift)
+      if not byte_val & 0x80:
+        break
+      shift += 7
+    return cls.FromZigZag(val)
+
+  def _AddValues(self, timestamp_value_pairs, offset=False,
+                 add_downsamples=True):
+    """Add values to this block.
+
+    Args:
+      timestamp_value_pairs: list of (unix_timestamp, value) tuples
+      offset: if True, values are offsets from previous value
+      add_downsamples: if True, add downsample values if necessary
+    """
+    if not timestamp_value_pairs:
+      return
+
+    timestamp_value_pairs.sort()
+
+    if offset:
+      # Convert to absolute values
+      prev_value = self.end_value
+      for pair in timestamp_value_pairs:
+        pair[1] += prev_value
+        prev_value = pair[1]
+
+    if add_downsamples and self.resolution == self.FULL:
+      self._CheckAddDownsamples(timestamp_value_pairs)
+
+    parts = [
+        self.times_and_values or '',
+    ]
+
+    for timestamp, value in timestamp_value_pairs:
+      if timestamp < self.start_time:
+        logging.warn('Skipping old value (%s): %d < %d',
+                     self.key().name(), timestamp, self.start_time)
+      parts.append(self.ToVarint(timestamp - self.end_time))
+      parts.append(self.ToVarint(value - self.end_value))
+      self.end_time = timestamp
+      self.end_value = value
+    self.times_and_values = ''.join(parts)
+
+    if len(self.times_and_values) > self._BLOB_LIMIT:
+      self._Split()
+
+  def _CheckAddDownsamples(self, timestamp_value_pairs):
+    """Add downsamples if necessary.
+
+    Args:
+      timestamp_value_pairs: sorted list of (unix_timestamp, value) tuples
+    """
+    # We add a skew factor so we don't have every timeseries trying to add
+    # downsamples at the same time.
+    skew = hash(self.parent_key().name())
+
+    assert self.resolution == self.FULL
+    for resolution, seconds in self.SECONDS.iteritems():
+      prev_timestamp = self.end_time
+      for timestamp, value in timestamp_value_pairs:
+        if (timestamp + skew) / seconds != (prev_timestamp + skew) / seconds:
+          # We've crossed a downsample boundary
+          Values.AddValue(self.parent_key(), timestamp, value,
+                          resolution=resolution)
+        prev_timestamp = timestamp
+
+  def GetValues(self):
+    """Fetch all values from this block.
+
+    Returns:
+      Sorted list of (unix_timestamp, value) tuples.
+    Raises:
+      InvalidSerializedData: serialized values stops in the middle of a
+        timestamp and value pair
+    """
+    values = [(self.start_time, self.start_value)]
+    if not self.times_and_values:
+      return values
+
+    last_timestamp = self.start_time
+    last_value = self.start_value
+
+    fh = StringIO.StringIO(self.times_and_values)
+    while True:
+      timestamp = self.FromVarint(fh)
+      if timestamp is None:
+        break
+      value = self.FromVarint(fh)
+      if value is None:
+        raise InvalidSerializedData('time without value')
+      last_timestamp += timestamp
+      last_value += value
+      values.append((last_timestamp, last_value))
+
+    values.sort(key=lambda x: x[0])
+
+    if self.resolution != self.FULL:
+      seconds = self.SECONDS[self.resolution]
+      # Clean up the values list for potential duplicates
+      i = 1
+      while i < len(values):
+        if values[i][0] / seconds == values[i - 1][0] / seconds:
+          # Latter value is a dupe
+          del values[i]
+        else:
+          i += 1
+
+    return values
+
+  def _Split(self, split_timestamp=None):
+    """Split this block into two.
+
+    Decodes the current block and puts the latter half of the values into a
+    new block.
+
+    Args:
+      split_timestamp: UNIX timestamp around which to split (inclusive in the
+        latter block)
+    """
+    values = self.GetValues()
+
+    if split_timestamp is None:
+      split_point = len(values) / 2
+    else:
+      split_point = bisect.bisect_left(values, (split_timestamp, 0))
+
+    logging.info(
+        'Splitting block from TimeSeries=%s, resolution=%d, '
+        'split_timestamp=%s, split_point=%d',
+        TimeSeries.KeyName(self.parent_key()), self.resolution,
+        split_timestamp, split_point)
+
+    old_values = values[:split_point]
+    if split_timestamp is None:
+      old_obj = self._Create(self.parent_key(),
+                             old_values,
+                             resolution=self.resolution,
+                             add_downsamples=False)
+      old_obj.save()
+    else:
+      # split_timestamp implies that we're doing an uneven block split. To
+      # avoid unnecessary fragmentation, add to the leading block in
+      # datastore, rather than creating a new one.
+      self._AddValuesDatastore(self.parent_key(),
+                               old_values,
+                               offset=False,
+                               resolution=self.resolution,
+                               add_downsamples=False)
+
+    self.times_and_values = None
+    self.start_time = values[split_point][0]
+    self.start_value = values[split_point][1]
+    self.end_time = self.start_time
+    self.end_value = self.start_value
+    self._AddValues(values[split_point + 1:], add_downsamples=False)
diff --git a/tools/collector.py b/tools/collector.py
new file mode 100644
index 0000000..61e003a
--- /dev/null
+++ b/tools/collector.py
@@ -0,0 +1,72 @@
+#!/usr/bin/python2.7
+
+import argparse
+import glob
+import socket
+import sys
+import time
+
+sys.path.append('..')
+
+from lib import tsdblib
+
+
+parser = argparse.ArgumentParser()
+parser.add_argument('--project', required=True)
+parser.add_argument('--base_url', required=True)
+FLAGS = parser.parse_args()
+
+
+class Collector(object):
+
+  def __init__(self):
+    self._start_time = int(time.time())
+    self._base_tags = self._BaseTags()
+    self._client = tsdblib.TSDBClient(FLAGS.base_url,
+                                      jit_callback=self._PutValues)
+
+  def _GetHostname(self):
+    return socket.gethostname()
+
+  def _GetMACAddresses(self):
+    for path in glob.iglob('/sys/class/net/*/address'):
+      mac_address = open(path, 'r').read().strip()
+      if mac_address == '00:00:00:00:00:00':
+        continue
+      yield mac_address
+
+  def _GetCollectorUptime(self):
+    return int(time.time()) - self._start_time
+
+  def _GetSystemUptime(self):
+    return int(open('/proc/uptime', 'r').read().split('.', 1)[0])
+
+  def _GetLoadAverage1m(self):
+    return int(float(open('/proc/loadavg', 'r').read().split(' ', 1)[0]) * 100)
+
+  def _BaseTags(self):
+    return [
+      ('project', FLAGS.project),
+      ('hostname', self._GetHostname()),
+    ] + [
+      ('mac_address', mac_address)
+      for mac_address in self._GetMACAddresses()
+    ]
+
+  def _CycleValues(self):
+    return [
+      ('collector_uptime_seconds', self._GetCollectorUptime()),
+      ('system_uptime_seconds', self._GetSystemUptime()),
+      ('system_load_average_1m', self._GetLoadAverage1m()),
+    ]
+
+  def _PutValues(self):
+    cycle_values = self._CycleValues()
+    for name, value in cycle_values:
+      self._client.PutValue(self._base_tags + [('value', name)], value)
+
+  def Loop(self):
+    time.sleep(9999999999)
+
+
+Collector().Loop()