geobed.go

package geobed

import (
	"archive/zip"
	"bufio"
	"bytes"
	"compress/gzip"
	"encoding/gob"
	geohash "github.com/TomiHiltunen/geohash-golang"
	"io"
	"log"
	"net/http"
	"os"
	"regexp"
	"sort"
	"strconv"
	"strings"
)

// There are over 2.4 million cities in the world. The Geonames data set only contains 143,270 and the MaxMind set contains 567,382 and 3,173,959 in the other MaxMind set.
// Obviously there's a lot of overlap and the worldcitiespop.txt from MaxMind contains a lot of dupes, though it by far the most comprehensive in terms of city - lat/lng.
// It may not be possible to have information for all cities, but many of the cities are also fairly remote and likely don't have internet access anyway.
// The Geonames data is preferred because it contains additional information such as elevation, population, and more. Population is good particuarly nice because a sense for
// the city size can be understood by applications. So showing all major cities is pretty easy. Though the primary goal of this package is to geocode, the additional information
// is bonus. So after checking the Geonames set, the geocoding functions will then look at MaxMind's.
// Maybe in the future this package will even use the Geonames premium data and have functions to look up nearest airports, etc.
// I would simply use just Geonames data, but there's so many more cities in the MaxMind set despite the lack of additional details.
//
// http://download.geonames.org/export/dump/cities1000.zip
// http://geolite.maxmind.com/download/geoip/database/GeoLiteCity_CSV/GeoLiteCity-latest.zip
// http://download.maxmind.com/download/worldcities/worldcitiespop.txt.gz

// A list of data sources.
var dataSetFiles = []map[string]string{
	{"url": "http://download.geonames.org/export/dump/cities1000.zip", "path": "./geobed-data/cities1000.zip", "id": "geonamesCities1000"},
	{"url": "http://download.geonames.org/export/dump/countryInfo.txt", "path": "./geobed-data/countryInfo.txt", "id": "geonamesCountryInfo"},
	{"url": "http://download.maxmind.com/download/worldcities/worldcitiespop.txt.gz", "path": "./geobed-data/worldcitiespop.txt.gz", "id": "maxmindWorldCities"},
	//{"url": "http://geolite.maxmind.com/download/geoip/database/GeoLiteCity_CSV/GeoLiteCity-latest.zip", "path": "./geobed-data/GeoLiteCity-latest.zip", "id": "maxmindLiteCity"},
}

// A handy map of US state codes to full names.
var UsSateCodes = map[string]string{
	"AL": "Alabama",
	"AK": "Alaska",
	"AZ": "Arizona",
	"AR": "Arkansas",
	"CA": "California",
	"CO": "Colorado",
	"CT": "Connecticut",
	"DE": "Delaware",
	"FL": "Florida",
	"GA": "Georgia",
	"HI": "Hawaii",
	"ID": "Idaho",
	"IL": "Illinois",
	"IN": "Indiana",
	"IA": "Iowa",
	"KS": "Kansas",
	"KY": "Kentucky",
	"LA": "Louisiana",
	"ME": "Maine",
	"MD": "Maryland",
	"MA": "Massachusetts",
	"MI": "Michigan",
	"MN": "Minnesota",
	"MS": "Mississippi",
	"MO": "Missouri",
	"MT": "Montana",
	"NE": "Nebraska",
	"NV": "Nevada",
	"NH": "New Hampshire",
	"NJ": "New Jersey",
	"NM": "New Mexico",
	"NY": "New York",
	"NC": "North Carolina",
	"ND": "North Dakota",
	"OH": "Ohio",
	"OK": "Oklahoma",
	"OR": "Oregon",
	"PA": "Pennsylvania",
	"RI": "Rhode Island",
	"SC": "South Carolina",
	"SD": "South Dakota",
	"TN": "Tennessee",
	"TX": "Texas",
	"UT": "Utah",
	"VT": "Vermont",
	"VA": "Virginia",
	"WA": "Washington",
	"WV": "West Virginia",
	"WI": "Wisconsin",
	"WY": "Wyoming",
	// Territories
	"AS": "American Samoa",
	"DC": "District of Columbia",
	"FM": "Federated States of Micronesia",
	"GU": "Guam",
	"MH": "Marshall Islands",
	"MP": "Northern Mariana Islands",
	"PW": "Palau",
	"PR": "Puerto Rico",
	"VI": "Virgin Islands",
	// Armed Forces (AE includes Europe, Africa, Canada, and the Middle East)
	"AA": "Armed Forces Americas",
	"AE": "Armed Forces Europe",
	"AP": "Armed Forces Pacific",
}

// Contains all of the city and country data. Cities are split into buckets by country to increase lookup speed when the country is known.
type GeoBed struct {
	c  Cities
	co []CountryInfo
}

type Cities []GeobedCity

func (c Cities) Len() int {
	return len(c)
}
func (c Cities) Swap(i, j int) {
	c[i], c[j] = c[j], c[i]
}
func (c Cities) Less(i, j int) bool {
	return toLower(c[i].City) < toLower(c[j].City)
}

// A combined city struct (the various data sets have different fields, this combines what's available and keeps things smaller).
type GeobedCity struct {
	City    string
	CityAlt string
	// TODO: Think about converting this to a small int to save on memory allocation. Lookup requests can have the strings converted to the same int if there are any matches.
	// This could make lookup more accurate, easier, and faster even. IF the int uses less bytes than the two letter code string.
	Country    string
	Region     string
	Latitude   float64
	Longitude  float64
	Population int32
	Geohash    string
}

// TODO: String interning? (much like converting country code to int)
// https://gist.github.com/karlseguin/6570372

// TODO: Store the cities in mmap...???
// https://github.com/boltdb/bolt/blob/master/bolt_unix.go#L42-L69
// Maybe even use bolt?

var maxMindCityDedupeIdx map[string][]string

// Holds information about the index ranges for city names (1st and 2nd characters) to help narrow down sets of the GeobedCity slice to scan when looking for a match.
var cityNameIdx map[string]int
var locationDedupeIdx map[string]bool

// Information about each country from Geonames including; ISO codes, FIPS, country capital, area (sq km), population, and more.
// Particularly useful for validating a location string contains a country name which can help the search process.
// Adding to this info, a slice of partial geohashes to help narrow down reverse geocoding lookups (maps to country buckets).
type CountryInfo struct {
	Country            string
	Capital            string
	Area               int32
	Population         int32
	GeonameId          int32
	ISONumeric         int16
	ISO                string
	ISO3               string
	Fips               string
	Continent          string
	Tld                string
	CurrencyCode       string
	CurrencyName       string
	Phone              string
	PostalCodeFormat   string
	PostalCodeRegex    string
	Languages          string
	Neighbours         string
	EquivalentFipsCode string
}

// Options when geocoding. For now just an exact match on city name, but there will be potentially other options that can be set to adjust how searching/matching works.
type GeocodeOptions struct {
	ExactCity bool
}

// An index range struct that's used for narrowing down ranges over the large Cities struct.
type r struct {
	f int
	t int
}

// Creates a new Geobed instance. You do not need more than one. You do not want more than one. There's a fair bit of data to load into memory.
func NewGeobed() GeoBed {
	g := GeoBed{}

	var err error
	g.c, err = loadGeobedCityData()
	g.co, err = loadGeobedCountryData()
	err = loadGeobedCityNameIdx()
	if err != nil || len(g.c) == 0 {
		g.downloadDataSets()
		g.loadDataSets()
		g.store()
	}

	return g
}

// Downloads the data sets if needed.
func (g *GeoBed) downloadDataSets() {
	os.Mkdir("./geobed-data", 0777)
	for _, f := range dataSetFiles {
		_, err := os.Stat(f["path"])
		if err != nil {
			if os.IsNotExist(err) {
				// log.Println(f["path"] + " does not exist, downloading...")
				out, oErr := os.Create(f["path"])
				defer out.Close()
				if oErr == nil {
					r, rErr := http.Get(f["url"])
					defer r.Body.Close()
					if rErr == nil {
						_, nErr := io.Copy(out, r.Body)
						if nErr != nil {
							// log.Println("Failed to copy data file, it will be tried again on next application start.")
							// remove file so another attempt can be made, should something fail
							err = os.Remove(f["path"])
						}
						r.Body.Close()
					}
					out.Close()
				} else {
					log.Println(oErr)
				}
			}
		}
	}
}

// Unzips the data sets and loads the data.
func (g *GeoBed) loadDataSets() {
	locationDedupeIdx = make(map[string]bool)

	for _, f := range dataSetFiles {
		// This one is zipped
		if f["id"] == "geonamesCities1000" {
			rz, err := zip.OpenReader(f["path"])
			if err != nil {
				log.Fatal(err)
			}
			defer rz.Close()

			for _, uF := range rz.File {
				fi, err := uF.Open()

				if err != nil {
					log.Fatal(err)
				}
				defer fi.Close()

				// Geonames uses a tab delineated format and it's not even consistent. No CSV reader that I've found for Go can understand this.
				// I'm not expecting any reader to either because it's an invalid CSV to be frank. However, we can still split up each row by \t
				scanner := bufio.NewScanner(fi)
				scanner.Split(bufio.ScanLines)

				i := 1
				for scanner.Scan() {
					i++

					// So regexp, sadly, must be used (well, unless I wanted parse each string byte by byte, pushing each into a buffer to append to a slice until a tab is reached, etc.).
					// But I'd have to also then put in a condition if the next byte was a \t rune, then append an empty string, etc. This just, for now, seems nicer (easier).
					// This is only an import/update, so it shouldn't be an issue for performance. If it is, then I'll look into other solutions.
					fields := regexp.MustCompile("\t").Split(scanner.Text(), 19)

					// NOTE: Now using a combined GeobedCity struct since not all data sets have the same fields.
					// Plus, the entire point was to geocode forward and reverse. Bonus information like elevation and such is just superfluous.
					// Leaving it here because it may be configurable... If options are passed to NewGeobed() then maybe Geobed can simply be a Geonames search.
					// Don't even load in MaxMind data...And if that's the case, maybe that bonus information is desired.
					if len(fields) == 19 {
						//id, _ := strconv.Atoi(fields[0])
						lat, _ := strconv.ParseFloat(fields[4], 64)
						lng, _ := strconv.ParseFloat(fields[5], 64)
						pop, _ := strconv.Atoi(fields[14])
						//elv, _ := strconv.Atoi(fields[15])
						//dem, _ := strconv.Atoi(fields[16])

						gh := geohash.Encode(lat, lng)
						// This is produced with empty lat/lng values - don't store it.
						if gh == "7zzzzzzzzzzz" {
							gh = ""
						}

						var c GeobedCity
						c.City = strings.Trim(string(fields[1]), " ")
						c.CityAlt = string(fields[3])
						c.Country = string(fields[8])
						c.Region = string(fields[10])
						c.Latitude = lat
						c.Longitude = lng
						c.Population = int32(pop)
						c.Geohash = gh

						// Don't include entries without a city name. If we want to geocode the centers of countries and states, then we can do that faster through other means.
						if len(c.City) > 0 {
							g.c = append(g.c, c)
						}
					}
				}
			}
		}

		// ...And this one is Gzipped (and this one may have worked with the CSV package, but parse it the same way as the others line by line)
		if f["id"] == "maxmindWorldCities" {
			// It also has a lot of dupes
			maxMindCityDedupeIdx = make(map[string][]string)

			fi, err := os.Open(f["path"])
			if err != nil {
				log.Println(err)
			}
			defer fi.Close()

			fz, err := gzip.NewReader(fi)
			if err != nil {
				log.Println(err)
			}
			defer fz.Close()

			scanner := bufio.NewScanner(fz)
			scanner.Split(bufio.ScanLines)

			i := 1
			for scanner.Scan() {
				i++
				t := scanner.Text()

				fields := strings.Split(t, ",")
				if len(fields) == 7 {
					var b bytes.Buffer
					b.WriteString(fields[0]) // country
					b.WriteString(fields[3]) // region
					b.WriteString(fields[1]) // city

					idx := b.String()
					b.Reset()
					maxMindCityDedupeIdx[idx] = fields
				}
			}

			// Loop the map of fields after dupes have been removed (about 1/5th less... 2.6m vs 3.1m inreases lookup performance).
			for _, fields := range maxMindCityDedupeIdx {
				if fields[0] != "" && fields[0] != "0" {
					if fields[2] != "AccentCity" {
						pop, _ := strconv.Atoi(fields[4])
						lat, _ := strconv.ParseFloat(fields[5], 64)
						lng, _ := strconv.ParseFloat(fields[6], 64)
						// MaxMind's data set is a bit dirty. I've seen city names surrounded by parenthesis in a few places.
						cn := strings.Trim(string(fields[2]), " ")
						cn = strings.Trim(cn, "( )")

						// Don't take any city names with erroneous punctuation either.
						if strings.Contains(cn, "!") || strings.Contains(cn, "@") {
							continue
						}

						gh := geohash.Encode(lat, lng)
						// This is produced with empty lat/lng values - don't store it.
						if gh == "7zzzzzzzzzzz" {
							gh = ""
						}

						// If the geohash was seen before...
						_, ok := locationDedupeIdx[gh]
						if !ok {
							locationDedupeIdx[gh] = true

							var c GeobedCity
							c.City = cn
							c.Country = toUpper(string(fields[0]))
							c.Region = string(fields[3])
							c.Latitude = lat
							c.Longitude = lng
							c.Population = int32(pop)
							c.Geohash = gh

							// Don't include entries without a city name. If we want to geocode the centers of countries and states, then we can do that faster through other means.
							if len(c.City) > 0 && len(c.Country) > 0 {
								g.c = append(g.c, c)
							}
						}
					}
				}
			}
			// Clear out the temrporary index (set to nil, it does get re-created) so that Go can garbage collect it at some point whenever it feels the need.
			maxMindCityDedupeIdx = nil
			locationDedupeIdx = nil
		}

		// ...And this one is just plain text
		if f["id"] == "geonamesCountryInfo" {
			fi, err := os.Open(f["path"])

			if err != nil {
				log.Fatal(err)
			}
			defer fi.Close()

			scanner := bufio.NewScanner(fi)
			scanner.Split(bufio.ScanLines)

			i := 1
			for scanner.Scan() {
				t := scanner.Text()
				// There are a bunch of lines in this file that are comments, they start with #
				if string(t[0]) != "#" {
					i++
					fields := regexp.MustCompile("\t").Split(t, 19)

					if len(fields) == 19 {
						if fields[0] != "" && fields[0] != "0" {
							isoNumeric, _ := strconv.Atoi(fields[2])
							area, _ := strconv.Atoi(fields[6])
							pop, _ := strconv.Atoi(fields[7])
							gid, _ := strconv.Atoi(fields[16])

							var ci CountryInfo
							ci.ISO = string(fields[0])
							ci.ISO3 = string(fields[1])
							ci.ISONumeric = int16(isoNumeric)
							ci.Fips = string(fields[3])
							ci.Country = string(fields[4])
							ci.Capital = string(fields[5])
							ci.Area = int32(area)
							ci.Population = int32(pop)
							ci.Continent = string(fields[8])
							ci.Tld = string(fields[9])
							ci.CurrencyCode = string(fields[10])
							ci.CurrencyName = string(fields[11])
							ci.Phone = string(fields[12])
							ci.PostalCodeFormat = string(fields[13])
							ci.PostalCodeRegex = string(fields[14])
							ci.Languages = string(fields[15])
							ci.GeonameId = int32(gid)
							ci.Neighbours = string(fields[17])
							ci.EquivalentFipsCode = string(fields[18])

							g.co = append(g.co, ci)
						}
					}
				}
			}
		}
	}

	// Sort []GeobedCity by city names to help with binary search (the City field is the most searched upon field and the matching names can be easily filtered down from there).
	sort.Sort(g.c)

	//debug
	//log.Println("TOTAL RECORDS:")
	//log.Println(len(g.c))

	// Index the locations of city names in the g.c []GeoCity slice. This way when searching the range can be limited so it will be faster.
	cityNameIdx = make(map[string]int)
	for k, v := range g.c {
		// Get the index key for the first character of the city name.
		ik := toLower(string(v.City[0]))
		if val, ok := cityNameIdx[ik]; ok {
			// If this key number is greater than what was previously recorded, then set it as the new indexed key.
			if val < k {
				cityNameIdx[ik] = k
			}
		} else {
			// If the index key has not yet been set for this value, then set it.
			cityNameIdx[ik] = k
		}

		// Get the index key for the first two characters of the city name.
		// if len(v.CityLower) >= 2 {
		// 	ik2 := v.CityLower[0:2]
		// 	if val, ok := cityNameIdx[ik2]; ok {
		// 		// If this key number is greater than what was previously recorded, then set it as the new indexed key.
		// 		if val < k {
		// 			cityNameIdx[ik2] = k
		// 		}
		// 	} else {
		// 		// If the index key has not yet been set for this value, then set it.
		// 		cityNameIdx[ik2] = k
		// 	}
		// }
	}
}

// Forward geocode, location string to lat/lng (returns a struct though)
func (g *GeoBed) Geocode(n string, opts ...GeocodeOptions) GeobedCity {
	var c GeobedCity
	n = strings.TrimSpace(n)
	if n == "" {
		return c
	}
	// variadic optional argument trick
	options := GeocodeOptions{}
	if len(opts) > 0 {
		options = opts[0]
	}

	if options.ExactCity {
		c = g.exactMatchCity(n)
	} else {
		// NOTE: The downside of this (currently) is that something is basically always returned. It's a best guess.
		// There's not much chance of it returning "not found" (or an empty GeobedCity struct).
		// If you'd rather have nothing returned if not found, look at more exact matching options.
		c = g.fuzzyMatchLocation(n)
	}

	return c
}

// Returns a GeobedCity only if there is an exact city name match. A stricter match, though if state or country are missing a guess will be made.
func (g *GeoBed) exactMatchCity(n string) GeobedCity {
	var c GeobedCity
	// Ignore the `abbrevSlice` value for now. Use `nCo` and `nSt` for more accuracy.
	nCo, nSt, _, nSlice := g.extractLocationPieces(n)
	nWithoutAbbrev := strings.Join(nSlice, " ")
	ranges := g.getSearchRange(nSlice)

	matchingCities := []GeobedCity{}

	// First, get everything that matches the city exactly (case insensitive).
	for _, rng := range ranges {
		// When adjusting the range, the keys become out of sync. Start from rng.f
		currentKey := rng.f
		for _, v := range g.c[rng.f:rng.t] {
			currentKey++
			// The full string (ie. "New York" or "Las Vegas")
			if strings.EqualFold(n, v.City) {
				matchingCities = append(matchingCities, v)
			}
			// The pieces with abbreviations removed
			if strings.EqualFold(nWithoutAbbrev, v.City) {
				matchingCities = append(matchingCities, v)
			}
			// Each piece - doesn't make sense for now. May revisit this.
			// ie. "New York" or "New" and "York" ... well, "York" is going to match a different city.
			// While that might be weeded out next, who knows. It's starting to get more fuzzy than I'd like for this function.
			// for _, np := range nSlice {
			// 	if strings.EqualFold(np, v.City) {
			// 		matchingCities = append(matchingCities, v)
			// 	}
			// }
		}
	}

	// If only one was found, we can stop right here.
	if len(matchingCities) == 1 {
		return matchingCities[0]
		// If more than one was found, we need to guess.
	} else if len(matchingCities) > 1 {
		// Then range over those matching cities and try to figure out which one it is - city names are unfortunately not unique of course.
		// There shouldn't be very many so I don't mind the multiple loops.
		for _, city := range matchingCities {
			// Was the state abbreviation present? That sounds promising.
			if strings.EqualFold(nSt, city.Region) {
				c = city
			}
		}

		for _, city := range matchingCities {
			// Matches the state and country? Likely the best scenario, I'd call it the best match.
			if strings.EqualFold(nSt, city.Region) && strings.EqualFold(nCo, city.Country) {
				c = city
			}
		}

		// If we still don't have a city, maybe we have a country with the city name, ie. "New York, USA"
		// This is tougher because there's a "New York" in Florida, Kentucky, and more. Let's use population to assist if we can.
		if c.City == "" {
			matchingCountryCities := []GeobedCity{}
			for _, city := range matchingCities {
				if strings.EqualFold(nCo, city.Country) {
					matchingCountryCities = append(matchingCountryCities, city)
				}
			}

			// If someone says, "New York, USA" they most likely mean New York, NY because it's the largest city.
			// Specific locations are often implied based on size or popularity even though the names aren't unique.
			biggestCity := GeobedCity{}
			for _, city := range matchingCountryCities {
				if city.Population > biggestCity.Population {
					biggestCity = city
				}
			}
			c = biggestCity
		}
	}

	return c
}

// When geocoding, this provides a scored best match.
func (g *GeoBed) fuzzyMatchLocation(n string) GeobedCity {
	nCo, nSt, abbrevSlice, nSlice := g.extractLocationPieces(n)
	// Take the reamining unclassified pieces (those not likely to be abbreviations) and get our search range.
	// These pieces are likely contain the city name. Narrowing down the search range will make the lookup faster.
	ranges := g.getSearchRange(nSlice)

	var bestMatchingKeys = map[int]int{}
	var bestMatchingKey = 0
	for _, rng := range ranges {
		// When adjusting the range, the keys become out of sync. Start from rng.f
		currentKey := rng.f

		for _, v := range g.c[rng.f:rng.t] {
			currentKey++

			// Mainly useful for strings like: "Austin, TX" or "Austin TX" (locations with US state codes). Smile if your location string is this simple.
			if nSt != "" {
				if strings.EqualFold(n, v.City) && strings.EqualFold(nSt, v.Region) {
					return v
				}
			}

			// Special case. Airport codes and other short 3 letter abbreviations, ie. NYC and SFO
			// Country codes could present problems here. It seems to work for NYC, but not SFO (which there are multiple SFOs actually).
			// Leaving it for now, but airport codes are tricky (though they are popular on Twitter). These must be exact (case sensitive) matches.
			// if len(n) == 3 {
			// 	alts := strings.Split(v.CityAlt, ",")
			// 	for _, altV := range alts {
			// 		if altV != "" {
			// 			if altV == n {
			// 				if val, ok := bestMatchingKeys[currentKey]; ok {
			// 					bestMatchingKeys[currentKey] = val + 4
			// 				} else {
			// 					bestMatchingKeys[currentKey] = 4
			// 				}
			// 			}
			// 		}
			// 	}
			// }

			// Abbreviations for state/country
			// Region (state/province)
			for _, av := range abbrevSlice {
				lowerAv := toLower(av)
				if len(av) == 2 && strings.EqualFold(v.Region, lowerAv) {
					if val, ok := bestMatchingKeys[currentKey]; ok {
						bestMatchingKeys[currentKey] = val + 5
					} else {
						bestMatchingKeys[currentKey] = 5
					}
				}

				// Country (worth 2 points if exact match)
				if len(av) == 2 && strings.EqualFold(v.Country, lowerAv) {
					if val, ok := bestMatchingKeys[currentKey]; ok {
						bestMatchingKeys[currentKey] = val + 3
					} else {
						bestMatchingKeys[currentKey] = 3
					}
				}
			}

			// A discovered country name converted into a country code
			if nCo != "" {
				if nCo == v.Country {
					if val, ok := bestMatchingKeys[currentKey]; ok {
						bestMatchingKeys[currentKey] = val + 4
					} else {
						bestMatchingKeys[currentKey] = 4
					}
				}
			}

			// A discovered state name converted into a region code
			if nSt != "" {
				if nSt == v.Region {
					if val, ok := bestMatchingKeys[currentKey]; ok {
						bestMatchingKeys[currentKey] = val + 4
					} else {
						bestMatchingKeys[currentKey] = 4
					}
				}
			}

			// If any alternate names can be discovered, take them into consideration.
			if v.CityAlt != "" {
				alts := strings.Fields(v.CityAlt)
				for _, altV := range alts {
					if strings.EqualFold(altV, n) {
						if val, ok := bestMatchingKeys[currentKey]; ok {
							bestMatchingKeys[currentKey] = val + 3
						} else {
							bestMatchingKeys[currentKey] = 3
						}
					}
					// Exact, a case-sensitive match means a lot.
					if altV == n {
						if val, ok := bestMatchingKeys[currentKey]; ok {
							bestMatchingKeys[currentKey] = val + 5
						} else {
							bestMatchingKeys[currentKey] = 5
						}
					}
				}
			}

			// Exact city name matches mean a lot.
			if strings.EqualFold(n, v.City) {
				if val, ok := bestMatchingKeys[currentKey]; ok {
					bestMatchingKeys[currentKey] = val + 7
				} else {
					bestMatchingKeys[currentKey] = 7
				}
			}

			for _, ns := range nSlice {
				ns = strings.TrimSuffix(ns, ",")

				// City (worth 2 points if contians part of string)
				if strings.Contains(toLower(v.City), toLower(ns)) {
					if val, ok := bestMatchingKeys[currentKey]; ok {
						bestMatchingKeys[currentKey] = val + 2
					} else {
						bestMatchingKeys[currentKey] = 2
					}
				}

				// If there's an exat match, maybe there was noise in the string so it could be the full city name, but unlikely. For example, "New" or "Los" is in many city names.
				// Still, give it a point because it could be the bulkier part of a city name (or the city name could be one word). This has helped in some cases.
				if strings.EqualFold(v.City, ns) {
					if val, ok := bestMatchingKeys[currentKey]; ok {
						bestMatchingKeys[currentKey] = val + 1
					} else {
						bestMatchingKeys[currentKey] = 1
					}
				}

			}
		}
	}

	// If no country was found, look at population as a factor. Is it obvious?
	if nCo == "" {
		hp := int32(0)
		hpk := 0
		for k, v := range bestMatchingKeys {
			// Add bonus point for having a population 1,000+
			if g.c[k].Population >= 1000 {
				bestMatchingKeys[k] = v + 1
			}
			// Now just add a bonus for having the highest population and points
			if g.c[k].Population > hp {
				hpk = k
				hp = g.c[k].Population
			}
		}
		// Add a point for having the highest population (if any of the results had population data available).
		if g.c[hpk].Population > 0 {
			bestMatchingKeys[hpk] = bestMatchingKeys[hpk] + 1
		}
	}

	m := 0
	for k, v := range bestMatchingKeys {
		if v > m {
			m = v
			bestMatchingKey = k
		}

		// If there is a tie breaker, use the city with the higher population (if known) because it's more likely to be what is meant.
		// For example, when people say "New York" they typically mean New York, NY...Though there are many New Yorks.
		if v == m {
			if g.c[k].Population > g.c[bestMatchingKey].Population {
				bestMatchingKey = k
			}
		}
	}

	// debug
	// log.Println("Possible results:")
	// log.Println(len(bestMatchingKeys))
	// for _, kv := range bestMatchingKeys {
	// 	log.Println(g.c[kv])
	// }
	// log.Println("Best match:")
	// log.Println(g.c[bestMatchingKey])
	// log.Println("Scored:")
	// log.Println(m)

	return g.c[bestMatchingKey]
}

// Splits a string up looking for potential abbreviations by matching against a shorter list of abbreviations.
// Returns country, state, a slice of strings with potential abbreviations (based on size; 2 or 3 characters), and then a slice of the remaning pieces.
// This does a good job at separating things that are clearly abbreviations from the city so that searching is faster and more accuarate.
func (g *GeoBed) extractLocationPieces(n string) (string, string, []string, []string) {
	var re = regexp.MustCompile("")

	// Extract all potential abbreviations.
	re = regexp.MustCompile(`[\S]{2,3}`)
	abbrevSlice := re.FindStringSubmatch(n)

	// Convert country to country code and pull it out. We'll use it as a secondary form of validation. Remove the code from the original query.
	nCo := ""
	for _, co := range g.co {
		re = regexp.MustCompile("(?i)^" + co.Country + ",?\\s|\\s" + co.Country + ",?\\s" + co.Country + "\\s$")
		if re.MatchString(n) {
			nCo = co.ISO
			// And remove it so we have a cleaner query string for a city.
			n = re.ReplaceAllString(n, "")
		}
	}

	// Find US State codes and pull them out as well (do not convert state names, they can also easily be city names).
	nSt := ""
	for sc, _ := range UsSateCodes {
		re = regexp.MustCompile("(?i)^" + sc + ",?\\s|\\s" + sc + ",?\\s|\\s" + sc + "$")
		if re.MatchString(n) {
			nSt = sc
			// And remove it too.
			n = re.ReplaceAllString(n, "")
		}
	}
	// Trim spaces and commas off the modified string.
	n = strings.Trim(n, " ,")

	// Now extract words (potential city names) into a slice. With this, the index will be referenced to pinpoint sections of the g.c []GeobedCity slice to scan.
	// This results in a much faster lookup. This is over a simple binary search with strings.Search() etc. because the city name may not be the first word.
	// This should not contain any known country code or US state codes.
	nSlice := strings.Split(n, " ")

	return nCo, nSt, abbrevSlice, nSlice
}

// There's potentially 2.7 million items to range though, let's see if we can reduce that by taking slices of the slice in alphabetical order.
func (g *GeoBed) getSearchRange(nSlice []string) []r {
	// NOTE: A simple binary search was not helping here since we aren't looking for one specific thing. We have multiple elements, city, state, country.
	// So we'd end up with multiple binary searches to piece together which could be quite a few exponentially given the possible combinations...And so it was slower.

	ranges := []r{}
	for _, ns := range nSlice {
		ns = strings.TrimSuffix(ns, ",")

		if len(ns) > 0 {
			// Get the first character in the string, this tells us where to stop.
			fc := toLower(string(ns[0]))
			// Get the previous index key (by getting the previous character in the alphabet) to figure out where to start.
			pik := string(prev(rune(fc[0])))

			// To/from key
			fk := 0
			tk := 0
			if val, ok := cityNameIdx[pik]; ok {
				fk = val
			}
			if val, ok := cityNameIdx[fc]; ok {
				tk = val
			}
			// Don't let the to key be out of range.
			if tk == 0 {
				tk = (len(g.c) - 1)
			}
			ranges = append(ranges, r{fk, tk})
		}
	}

	return ranges
}

func prev(r rune) rune {
	return r - 1
}

// Reverse geocode
func (g *GeoBed) ReverseGeocode(lat float64, lng float64) GeobedCity {
	c := GeobedCity{}

	gh := geohash.Encode(lat, lng)
	// This is produced with empty lat/lng values - don't look for anything.
	if gh == "7zzzzzzzzzzz" {
		return c
	}

	// Note: All geohashes are going to be 12 characters long. Even if the precision on the lat/lng isn't great. The geohash package will center things.
	// Obviously lat/lng like 37, -122 is a guess. That's no where near the resolution of a city. Though we're going to allow guesses.
	mostMatched := 0
	matched := 0
	for k, v := range g.c {
		// check first two characters to reduce the number of loops
		if v.Geohash[0] == gh[0] && v.Geohash[1] == gh[1] {
			matched = 2
			for i := 2; i <= len(gh); i++ {
				//log.Println(gh[0:i])
				if v.Geohash[0:i] == gh[0:i] {
					matched++
				}
			}
			// tie breakers go to city with larger population (NOTE: There's still a chance that the next pass will uncover a better match)
			if matched == mostMatched && g.c[k].Population > c.Population {
				c = g.c[k]
				// log.Println("MATCHES")
				// log.Println(matched)
				// log.Println("CITY")
				// log.Println(c.City)
				// log.Println("POPULATION")
				// log.Println(c.Population)
			}
			if matched > mostMatched {
				c = g.c[k]
				mostMatched = matched
			}
		}
	}

	return c
}

// A slightly faster lowercase function.
func toLower(s string) string {
	b := make([]byte, len(s))
	for i := range b {
		c := s[i]
		if c >= 'A' && c <= 'Z' {
			c += 'a' - 'A'
		}
		b[i] = c
	}
	return string(b)
}

// A slightly faster uppercase function.
func toUpper(s string) string {
	b := make([]byte, len(s))
	for i := range b {
		c := s[i]
		if c >= 'a' && c <= 'z' {
			c -= 'a' - 'A'
		}
		b[i] = c
	}
	return string(b)
}

// Dumps the Geobed data to disk. This speeds up startup time on subsequent runs (or if calling NewGeobed() multiple times which should be avoided if possible).
// TODO: Refactor
func (g GeoBed) store() error {
	b := new(bytes.Buffer)

	// Store the city info
	enc := gob.NewEncoder(b)
	err := enc.Encode(g.c)
	if err != nil {
		b.Reset()
		return err
	}

	fh, eopen := os.OpenFile("./geobed-data/g.c.dmp", os.O_CREATE|os.O_WRONLY, 0666)
	defer fh.Close()
	if eopen != nil {
		b.Reset()
		return eopen
	}
	n, e := fh.Write(b.Bytes())
	if e != nil {
		b.Reset()
		return e
	}
	log.Printf("%d bytes successfully written to cache file\n", n)

	// Store the country info as well (this is all now repetition - refactor)
	b.Reset()
	//enc = gob.NewEncoder(b)
	err = enc.Encode(g.co)
	if err != nil {
		b.Reset()
		return err
	}

	fh, eopen = os.OpenFile("./geobed-data/g.co.dmp", os.O_CREATE|os.O_WRONLY, 0666)
	defer fh.Close()
	if eopen != nil {
		b.Reset()
		return eopen
	}
	n, e = fh.Write(b.Bytes())
	if e != nil {
		b.Reset()
		return e
	}
	log.Printf("%d bytes successfully written to cache file\n", n)

	// Store the index info (again there's some repetition here)
	b.Reset()
	//enc = gob.NewEncoder(b)
	err = enc.Encode(cityNameIdx)
	if err != nil {
		b.Reset()
		return err
	}

	fh, eopen = os.OpenFile("./geobed-data/cityNameIdx.dmp", os.O_CREATE|os.O_WRONLY, 0666)
	defer fh.Close()
	if eopen != nil {
		b.Reset()
		return eopen
	}
	n, e = fh.Write(b.Bytes())
	if e != nil {
		b.Reset()
		return e
	}
	log.Printf("%d bytes successfully written to cache file\n", n)

	b.Reset()
	return nil
}

// Loads a GeobedCity dump, which saves a bit of time.
func loadGeobedCityData() ([]GeobedCity, error) {
	fh, err := os.Open("./geobed-data/g.c.dmp")
	if err != nil {
		return nil, err
	}
	gc := []GeobedCity{}
	dec := gob.NewDecoder(fh)
	err = dec.Decode(&gc)
	if err != nil {
		return nil, err
	}
	return gc, nil
}

func loadGeobedCountryData() ([]CountryInfo, error) {
	fh, err := os.Open("./geobed-data/g.co.dmp")
	if err != nil {
		return nil, err
	}
	co := []CountryInfo{}
	dec := gob.NewDecoder(fh)
	err = dec.Decode(&co)
	if err != nil {
		return nil, err
	}
	return co, nil
}

func loadGeobedCityNameIdx() error {
	fh, err := os.Open("./geobed-data/cityNameIdx.dmp")
	if err != nil {
		return err
	}
	dec := gob.NewDecoder(fh)
	cityNameIdx = make(map[string]int)
	err = dec.Decode(&cityNameIdx)
	if err != nil {
		return err
	}
	return nil
}